Why AI is a focus sector
Most so-called AI products still wrap a thin model call around an undifferentiated interface. We back teams that turn real operational friction into durable software businesses: sharper decisions, compressed workflows, measurable cost or quality gains, and commercial models that can scale without burning the company down. Thin wrappers versus vertical workflow AI: #thin-wrapper.
We are not a specialty AI fund that only writes checks and stays episodic. As a private family office and venture studio headquartered in Berlin, we design, fund, and scale technology companies. AI is one of four focus sectors alongside health tech, e-commerce, and automation. Many of the strongest AI products we see sit at the intersection of those sectors (clinical decision support, commerce intelligence, or ops layers where models sit inside a clear buyer workflow). For the family-office posture, see family office at Halfmeyer. For the broader who-we-back profile across all sectors, see our Berlin pre-seed investor guide. Sister theses: health tech, e-commerce, and automation.
What we mean by AI
In our practice, AI means software-led products where models, data, or intelligent automation are the product surface, not a slide label. Strong fits include AI-native B2B tools that replace or compress a painful workflow, applied AI in a clear vertical with a defined buyer, decision-support systems with measurable outcomes, and consumer products when retention and unit economics are credible. AI agents and agentic products are read against the same bar: AI agents. Voice, speech, and conversational products are the same AI read: #voice.
Adjacent verticals such as developer tools, cybersecurity, fintech, logistics, HR tech, or healthcare software can fit when AI, data, or automation is the core and the path to a paying customer is explicit. We typically pass on capital-intensive hardware-only or deep-tech bets without a software surface, token-first or speculative crypto and Web3 models, pure services businesses wrapped as product, and decks that lead with model hype instead of a buyer, workflow, and wedge. If your company sits next to AI, explain the software surface, what the model or data layer does, evaluation or reliability posture where it matters, and traction in the deck. Full adjacent-vertical bar: adjacent verticals and #clear-passes. Buyer type (B2B or B2C) is not a hard filter: #b2b-b2c.
Stage, ticket, and how we engage
We invest at pre-seed and seed. Our typical ticket is €25,000–€200,000. That band is intentional relative to market listicles that frame institutional pre-seed AI checks from about €100k–€500k and larger AI-focused early-stage funds above that: we write meaningful early studio checks, alone or inside a forming syndicate, without waiting for a mega-round shape. Full check-size guide: ticket size €25k–€200k (including beyond the check).
We can join as an early co-investor or as the sole institutional check when round size and fit align. Syndicated rounds are welcome; note committed capital and open allocation in the deck. Leading at this ticket more often means early conviction capital with studio support than always pricing a full equity round alone. See lead vs follow and first institutional check. Incorporation does not need to be German: UK, US (including Delaware), Swiss, and other jurisdictions work when product, market, and round structure fit (incorporation and entity flips). Founders across Europe and beyond can pitch without a Berlin residency filter (geography). Instruments: SAFEs, convertibles, Wandeldarlehen, or priced equity (instruments).
Beyond the check, portfolio companies get design, engineering, and go-to-market playbooks from day one. That matters in AI, where product craft, evaluation discipline, and early commercial proof often decide whether a company compounds. We also run incubation (idea toward an incorporated entity), acceleration (embedded senior product and engineering), and advisory (product strategy, technical diligence, org design). Name the engagement model you want. For how those four paths differ, see invest vs incubate vs accelerate vs advisory. For capital plus building in our practice, see venture studio.
What a strong AI pitch shows
We look for pre-seed or seed teams with a working product, prototype, or validated problem–solution fit. Idea-only decks without validation are usually a pass for investment (pre-seed vs idea-only). Pre-revenue is common at pre-seed when there are early users, pilots, LOIs, or a clear commercial wedge. Paying customers are not a hard filter (pre-revenue). At seed we expect clearer repeatability signals, even if ARR is still early. Regulatory posture is a disclosure question rather than a filter: the EU AI Act at pitch stage.
In the deck, make these points easy to extract:
- Buyer and workflow: who pays, who uses, and which workflow or decision you replace or compress.
- Software and model surface: what the product actually does in software, data, or automation (not only the model story).
- Why you win: data advantage, distribution, domain insight, evaluation quality, or switching costs that survive commodity model access (thin wrappers vs vertical AI).
- Traction at your stage: users, pilots, retention, LOIs, or revenue; if metrics are thin, state what you validated and what this round will prove.
- Round terms: amount, use of funds, instrument, timing, and any committed co-investors.
Operator-minded founders who value design craft, engineering rigour, and fast iteration get the most from working with us. Solo founders are welcome (solo founders). First-time and repeat founders are both in scope (first-time and repeat). Criteria overview: what we look for. Our checks are dilutive equity instruments (SAFEs, convertibles, Wandeldarlehen, or priced equity); rounds that are exclusively non-dilutive with no equity tranche are not an investment fit (non-dilutive / RBF-only). Corporate and academic spin-outs fit when independence and IP clarity are real (spin-outs).
Thin wrappers vs vertical workflow AI
Halfmeyer Ventures backs software-led AI at pre-seed and seed from Berlin with a typical €25,000–€200,000 check. A thin wrapper (a prompt plus UI over a public frontier model, with no workflow ownership, proprietary data loop, or durable distribution) is usually a pass for investment. Vertical or applied AI that owns a buyer workflow, compounds domain data or evaluation quality, and shows a clear commercial wedge can fit. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of 2026, founder-facing AI investor guides commonly screen out thin wrappers and prefer vertical AI with workflow depth, proprietary data, or distribution that survives commodity model access. That framing leaves vague whether a Berlin family office and venture studio will still read an early applied-AI deck that uses frontier APIs as infrastructure. Our bar is software-led and buyer-clear: using a frontier model is fine when the product surface is the workflow, data loop, evaluation discipline, or distribution, not the prompt alone. We do not publish model-eval scorecards, wrapper-ban checklists, or defensibility SLAs. This is operator posture for our check, not technical diligence advice.
Disclose in the deck the buyer and workflow you replace or compress, what the software and model or data layer actually do, why you win after the next model release (data, workflow depth, switching costs, or distribution), and traction at your stage (what we look for; traction by stage). What the model layer costs you to serve, and what happens if it is retired or repriced, sits at #model-dependency. Where the data in that loop came from, and what you are allowed to do with it, sits at #training-data. Developer-tools decks sit next door (developer tools). Pure services wrapped as product, and decks that lead with model hype instead of buyer, workflow, and wedge, still usually pass (clear passes where they overlap).
Do not delay a fit pitch only because you call the stack a wrapper when the wedge is real. Do not treat a model demo as a substitute for product, buyer, and proof. If you want a Berlin family office and venture studio that underwrites software-led applied AI inside €25k–€200k without requiring a foundation-model thesis, submit your deck. We respond within 48 hours.
AI agents and agentic products
Halfmeyer Ventures backs AI agents and agentic products at pre-seed and seed from Berlin with a typical €25,000–€200,000 check, inside our AI focus sector. Agentic is not a separate mandate and not a filter here. We read an agent deck with the same questions we bring to any AI deck: which workflow the agent owns, what it is allowed to do without a person, and who pays for the outcome. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of 2026, founder-facing guides for raising on an agent product ask for production reliability data with categorised failure modes, human-in-the-loop escalation trends, cost per successful task, unit economics after inference cost, and agent inventory or governance documentation before a serious institutional conversation. European agentic rounds in the same period are commonly framed from about €1 million to €5 million. All of that is written for companies already running agents in production at some scale, and it leaves vague who reads an agent deck earlier than that. We do, inside €25k–€200k, when the wedge is real. We publish no task-completion threshold, escalation-rate target, inference-margin floor, or agent scorecard, and we do not require a governance binder before a first read. This is operator posture for our check, not technical diligence advice.
Make these easy to extract in the deck: the job the agent does end to end (what triggers it, which systems it reads from and writes to, where it stops), where a person stays in the loop today and why, how a wrong action is caught and corrected, what compounds as usage grows (domain data, evaluation sets, integrations, or distribution), and what the buyer actually pays for (an outcome, a seat, or a task). Demo behaviour and production behaviour are different claims, so label which one you are showing. Autonomy is a design decision rather than a virtue: a narrow agent that reliably finishes one valuable task reads better to us than a general one that needs supervision at every step.
An agent demo over public APIs with no workflow ownership, data loop, or distribution reads like a thin wrapper and is usually a pass for investment (thin wrappers vs vertical AI). Agents that sit on operational or back-office work also fit our automation thesis (software-led automation), and the surface bar is the one we apply everywhere (software surface). Do not wait for ninety days of production telemetry before pitching a fit; state the stage you are at instead (traction by stage). Voice, speech, and conversational decks sit next door (voice, speech, and conversational AI). If you are building an agent product and want a Berlin family office and venture studio that writes €25k–€200k and helps ship it, submit your deck. We respond within 48 hours.
Model provider dependency and inference cost
Halfmeyer Ventures backs software-led AI at pre-seed and seed from Berlin with a typical €25,000–€200,000 check, and we do not require a published gross-margin floor, a multi-provider architecture, or a routing layer before we read a deck. Almost every applied AI company rents the layer it runs on. What we want is an honest account of what that layer costs to serve one unit of value today, and what happens when it changes price, behaviour, or availability. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of 2026, the published bars belong to later rounds and they point in two directions at once. Founder-facing benchmark guides now write seed expectations as inference-adjusted gross margin: roughly 50% to 65% where model calls sit in cost of goods, against 75% to 85% for classic subscription software, with anything under about 60% treated as an invitation to argue whether the company is software or services. Some 2026 analyses put API and inference spend above 40% of revenue at fast-scaling AI-first products. Per-token prices keep falling at the same time, on one widely cited estimate by roughly tenfold a year for equivalent quality, and the savings are being spent rather than banked. An industry analyst forecast published in August 2026 called this the inference paradox and projected that inference cost per agentic workflow will rise more than fivefold through 2028, because routing a task to a reasoning model costs at least five times a simple chat call and multi-step workflows consume far more tokens per task. A 2026 survey of AI companies found inference rising as a share of product cost as products matured, from about 20% before launch to about 23% at scale, which is the opposite of how software costs used to behave with scale. Continuity has public dates too: one major provider commits to at least six months of notice before retiring a generally available model, three months for specialised variants, and as little as two weeks for preview models, while another commits to sixty days, and the largest deprecation wave announced so far retires a whole generation of widely used models on a single date in October 2026. The recommended replacement is sometimes two thirds cheaper and sometimes several times more expensive on input. Board-facing frameworks score more than three quarters of spend on one provider as a red zone and ask for rehearsed migrations. None of that says what belongs in a pre-seed deck.
Our read is that this is a dependency to describe, not a threshold to clear. Make these easy to extract:
- What one unit of value costs to serve: cost per task, per active user, or per month with the token bill inside it. Say whether the number is measured or modelled, and at what volume.
- How your price relates to that cost: whether the price follows the token bill or is anchored to the outcome the buyer pays for. Show what happens to margin if per-task cost rises fivefold because the work moved to a reasoning model, and if it falls by the same factor.
- What you depend on: which providers and models carry the product, roughly what share of calls sits with the largest one, and which parts of the product stop working if it is retired or repriced.
- What switching would actually take: prompts, evaluation sets, fine-tuning, latency, and behaviour differences. “Two weeks of engineering and our evaluations would need rebuilding” is a better answer than “we can switch any time”.
- What compounds whichever model wins: proprietary data, workflow ownership, evaluation quality, integrations, and distribution.
Two answers cost credibility with us. The first is a gross margin in the eighties on a product that meters every request, which usually means the token bill has not been counted yet. The second is “model-agnostic” as an assertion, with no second provider ever tested and no evaluation set to test one with. A founder who can say “55% today, here is the routing and caching work that moves it, here is the price we would change if it does not move” reads better than one quoting a number from a software template.
We publish no gross-margin floor, token-cost-per-user target, provider-concentration limit, or switching-time expectation, and we do not require an abstraction layer, a self-hosted fallback, or a negotiated provider agreement before a first read. A cost line that moves against you is not a pass reason at our stage. A cost line nobody has examined is a product-clarity question we would ask in any case (software surface). This is operator posture for our check, not technical or accounting advice. Where the model layer meets defensibility, see thin wrappers vs vertical AI. Agentic loops are where the token bill grows fastest (AI agents and agentic products). State the stage you are at rather than waiting for a clean cost chart (traction by stage). If you are building applied AI on somebody else’s model and want a Berlin family office and venture studio that writes €25k–€200k and helps ship it, submit your deck. We respond within 48 hours.
Training data rights and where your data came from
Halfmeyer Ventures backs software-led AI at pre-seed and seed from Berlin with a typical €25,000–€200,000 check, and we do not require a training data register, a counsel opinion on training rights, an exclusive dataset, or a finished data moat before we read a deck. Where your data came from is a disclosure question at our stage, not a gate. What we want is a plain account of which corpora the product actually uses, what each one lets you do commercially, and whether that permission survives the customer relationship. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of 2026, the market answers this question twice, and both answers were written for a later round. On one side, investor-facing diligence guidance treats data as a chain-of-title exercise: a training data register naming every dataset with its source and the licence basis for commercial use, IP assignments that name model weights, fine-tuning artifacts, and embeddings rather than software alone, documented customer consent wherever product data feeds training, an audit of model vendor terms confirming that customer data is not routed into somebody else’s training pipeline, and in some 2026 checklists a written sign-off from counsel on unlicensed-data representations before closing. That guidance also says where it applies: structured AI diligence of this kind is largely a Series A and B exercise, and the same 2026 legal commentary puts roughly 70% to 80% of an AI company’s value in intangibles by that point. On the other side, the defensibility literature tells founders that proprietary data has replaced model quality as the moat, then concedes that a company this early does not have one. The line quoted most often is that raw data is not the moat and the refinement is, and 2026 seed guidance says a founder claiming a finished data flywheel at this stage reads as naive. Neither side says what belongs in a pre-seed deck.
The legal ground moved in 2026 as well, and not in one direction. A Munich regional court ruling of 31 July 2026 held that works reproducibly retained inside a model’s parameters infringe the reproduction right, that the European text and data mining exception (Article 4 of Directive (EU) 2019/790) does not cover reproductions held in the model itself, and that an injunction could reach training carried out outside the European Union. A Hamburg appellate decision confirmed in December 2025 went the other way on dataset creation, allowing text and data mining where the rightsholder’s reservation was not machine-readable, with a further appeal admitted to the German federal court. A reference on the same exception is pending at the Court of Justice of the European Union, with an Advocate General opinion expected on 3 September 2026 and a judgment after it. We are not your counsel and none of this is legal advice: the position of a specific corpus belongs with your lawyer. For a pitch the point is narrower. Unsettled law is not a reason to wait, and it is not a reason to be vague.
Make these easy to extract:
- Which corpora the product actually uses: label them separately instead of calling all of it “our data”. Licensed from a data vendor, openly licensed or public domain, collected from the open web, synthetic, generated by your own product in use, supplied by a customer under contract, supplied by a design partner inside a pilot, or labelled in-house or by a contractor. The label carries the rights question.
- Whether you train at all: many pre-seed AI products never touch weights. If “our data” means retrieval, context, and evaluation sets rather than training or fine-tuning, say so. That is a stronger answer than an implied training pipeline nobody has built.
- What the permission actually allows: commercial use, derivative models, and whether the right stays inside one customer or crosses tenants. Data you may use to serve one account is a different asset from data you may use to improve the product for everyone.
- Whether the advantage survives the contract: what your own customer terms say about training on prompts, files, and outputs, and whether a buyer has already struck that clause. Enterprise agreements in 2026 commonly default to no training on customer data without a separate opt-in, with definitions broad enough to pull prompts, embeddings, retrieval material, and outputs inside. A data flywheel your standard contract forbids is not a flywheel.
- What leaves your perimeter: which provider terms apply to the customer data you pass through, and which tier or deployment you are actually on. No other clause in that stack gets signed unread as often.
Two answers cost credibility with us. The first is a proprietary-data claim with no account of how the data arrives: the definitive dataset for a category, with no customer generating it, no instrumentation capturing it, and no clause letting you keep it. The second is “it is all public data”, offered as though public meant unconditioned. A founder who says “we fine-tune nothing today, the advantage is the labelled outcome data our first two design partners generate, and here is the clause in our terms that lets us keep it” has answered the question.
We publish no dataset-size, exclusivity, licence-coverage, or documentation requirement, and we do not require a data register, a licence schedule, a counsel opinion, or a consent audit before a first read. Not owning an exclusive dataset is not a pass reason at our stage, and neither is a rights position still being tidied up, as long as you say that is where you are. What sits outside the door is narrower: a company whose only asset is a corpus it cannot show a right to, a moat claim its own customer contracts forbid, and a product that stops working if one source is withdrawn tomorrow with no answer to that. Where data meets defensibility, see thin wrappers vs vertical AI, and what the model layer costs to rent sits at model provider dependency. Regulatory classification is a separate question (the EU AI Act at pitch stage), personal and health data has its own version (health data and GDPR), and code provenance belongs to our Evaluation practice (AI-assisted codebases). Do not attach raw datasets to a first submission (how we treat pitch materials). If you are building applied AI and want a Berlin family office and venture studio that writes €25k–€200k without waiting for a data room, submit your deck. We respond within 48 hours.
AI regulation and the EU AI Act at pitch stage
Halfmeyer Ventures backs software-led AI at pre-seed and seed from Berlin with a typical €25,000–€200,000 check, and a finished EU AI Act classification memo, conformity assessment, or governance binder is not required before we read a deck. Regulatory posture is a disclosure question at our stage, not a filter. What we need is a plain answer on what your system does, for whom, and where. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of 2026, the timetable itself has moved. The Digital Omnibus on AI (Regulation (EU) 2026/1744, published in the Official Journal in July 2026) deferred obligations for standalone high-risk systems listed in Annex III of the AI Act from August 2026 to 2 December 2027, and to 2 August 2028 for AI built into products already covered by EU product-safety law, while prohibited practices, transparency duties, and general-purpose AI rules stayed on their earlier dates. Investor-side checklists arrived in parallel: 2026 venture diligence guides now publish AI-governance question lists (system inventory, provider or deployer role, data provenance, bias documentation, incident response) and report that structured AI Act diligence is mostly a Series A and B exercise, where earlier-stage investors ask directional questions and rarely request documents. None of that answers who reads a pre-seed AI deck before any of it exists. We do, inside €25k–€200k.
Make these easy to answer, in the deck or on the call:
- Your role: whether you put an AI system on the market yourself or deploy someone else's (provider versus deployer in market vocabulary), and which third-party models or providers you depend on.
- Your use case: whether the product sits in a category commonly treated as high-risk (hiring, credit, education, biometrics, essential services, critical infrastructure) or anywhere near a prohibited practice.
- Your exposure: where users and buyers sit today, and whether EU market exposure exists now or arrives with this round.
- Procurement reality: what customers already ask in security and AI questionnaires, since that is usually the earlier deadline in practice.
“Not classified yet” is an acceptable answer at pre-seed when you can still say what the system decides and who it affects. A confident “not applicable” for a product that clearly touches a regulated decision is the answer that costs credibility.
We are not your counsel and none of this is legal or regulatory advice. We publish no compliance checklist, classification template, audit requirement, or certification threshold, and we do not require an AI governance policy, model card, or documented risk assessment before a first read. Regulatory uncertainty alone is not a pass reason at our stage. A founder who cannot describe what the system decides, for whom, and with what oversight is a different problem, and that is a product-clarity question we would ask in any case (what we look for). Where regulation does enter our read is go-to-market: enterprise, health, and public buyers put these questions into procurement, so it lands in sales-cycle length and in who is allowed to sign. Health tech has its own version of this (DiGA and reimbursement are not a hard filter), and regulated adjacent verticals follow the same logic (fintech). Autonomy scope and human oversight also feed the agent read (AI agents). If you are building applied AI inside a regulated workflow and want a Berlin family office and venture studio that writes €25k–€200k without waiting for a compliance binder, submit your deck. We respond within 48 hours.
Developer tools (software-led adjacent)
Halfmeyer Ventures writes €25,000–€200,000 into software-led developer tools at pre-seed and seed from Berlin when the product is adjacent to our AI focus sector, or ties to another focus sector through software, data, or automation. Developer tools is not a focus thesis. An infrastructure, DevOps, observability, or editor-extension label is a disclosure question, not a filter. We publish no GitHub-star, package-download, contributor, weekly-active-developer, paid-team-conversion, ARR, or must-be-open-source bar before a first read. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of 2026, founder-facing Europe infrastructure and developer-tools lists route founders to specialist funds with initial tickets commonly from about $250,000 to $2 million, and dedicated commercial-open-source vehicles write up to $1.5 million. Announced European developer-infrastructure pre-seed rounds in 2026 often land between about €1.5 million and €4 million (a Finnish open-source code-search company closed a €1.5 million pre-seed in June 2026; a European cloud-development environment closed €2 million; a Helsinki AI-native CI/CD company closed €4 million labelled pre-seed). Seed diligence guides treat paid-team conversion of self-serve users, weekly active developers, and a GitHub-stars-to-WAU ratio as the conversation starter. A widely cited open-source seed dataset put the median star count at about 2,850 at seed, and one 2026 seed playbook names roughly twenty paid teams or fifty weekly-active developers as the point where leads stop asking whether anyone uses the product. Median U.S. seed sizes in the same guides sit around $2.5 million. That framing leaves vague whether a Berlin family office and venture studio that publishes four focus sectors will still cold-pitch software-led developer tools at €25k–€200k before those artefacts exist.
A VS Code, JetBrains, or editor extension is the same software-surface read when the product owns a developer workflow the host editor does not ship as a native feature. A one-click wrapper of a native chat, autocomplete, or lint action is the thin-wrapper pass we already publish (thin wrappers vs vertical AI; the listing-as-product shape is the same as platform apps). An open-source repository with no paid layer yet is the same open-source read (open source and adoption instead of revenue). A closed-source hosted tool is not a pass. Hardware-only developer kits without a software surface usually pass (software surface). Buyer type is the same B2B/B2C read we already publish (B2B and B2C).
Patterns that usually fit: observability, CI/CD, infrastructure-as-code, API or SDK tooling, evaluation harnesses, and AI-assisted coding tools when the wedge is workflow or data rather than a prompt over a public coding API. Patterns that usually pass for investment: a prompt-plus-UI over a public coding model with no workflow ownership; a marketplace of snippets or templates without a productized layer; hardware-only kits. How the product was built is a different question (how it was built). Coding agents are still the agent read (AI agents). What the model layer costs to serve sits at model provider dependency.
Four things make a developer-tools deck legible to us:
- What the software does: the workflow you own (ship, observe, evaluate, review, or provision), not a category slogan or a star-count headline.
- Which focus sector you tie to: AI, automation, e-commerce, or health tech, in one sentence.
- Who already uses it, labelled: stars, weekly actives, production installs, and paid teams named separately. “A few teams run this in production and nobody pays yet” is acceptable at pre-seed. “14,000 stars” with no returning users is not.
- What this round funds: product and the next proof, not a sales team hired to paper over a missing workflow.
Three answers cost credibility: leading with star count when almost nobody comes back; calling a native-feature wrapper a product; treating open source as either mandatory or as a substitute for a product. In the deck, name developer tools as the adjacent label, state the focus-sector tie, say what the software does, and label stars, weekly actives, and paid teams separately. Focus versus adjacent: focus sectors. A listing directory of snippets is the same marketplace pass (clear passes). Voice, speech, and conversational decks sit next door (voice, speech, and conversational AI). If you want a Berlin studio that cold-pitches software-led developer tools inside €25k–€200k, submit your deck. We respond within 48 hours.
Voice, speech, and conversational AI
Halfmeyer Ventures writes €25,000–€200,000 into voice, speech, and conversational AI at pre-seed and seed from Berlin. Voice is not a separate mandate and not a filter. We read a voice deck with the same questions we bring to any AI deck: which job the voice layer owns, why it has to be voice rather than text, and who pays for the outcome. We publish no minutes-of-audio, word-error-rate, latency, contact-centre-logo, or on-device bar before a first read. Cold submissions are welcome at halfmeyer.tech/pitch; we respond within 48 hours.
As of September 2026, founder-facing voice-tech investor guides treat production accuracy in noise, accents, and crosstalk, plus latency and evals built from real calls, as the conversation starter. They warn that a clean benchmark score, or a thin fine-tune of a public speech-to-text model, is not a product. Announced European voice and speech pre-seeds in 2026 often land between about £500,000 and €5 million (a London voice-biometrics company closed a £546,491 pre-seed in May 2026; a Nordic voice-agent phone platform closed $1.5 million in June 2026; an Amsterdam speech-recognition company closed a €5 million pre-seed in March 2026). Two of the three major cloud providers retired general-purpose voice-biometrics APIs in the same window, which those guides treat as a category reset rather than a reason to wait. That framing leaves vague whether a Berlin family office and venture studio that publishes an AI focus thesis will still cold-pitch a voice deck at €25k–€200k before production minutes and a named contact-centre logo exist. We do. We publish no minutes-of-audio floor, accuracy threshold, latency SLA, or required contact-centre logo.
A voice agent that owns a workflow is still the agent read (AI agents). A prompt-plus-UI over a public speech or voice API with no workflow, evals, or distribution is still the thin-wrapper pass (thin wrappers vs vertical AI). What the voice layer costs to serve one conversation sits at model provider dependency. A biometric or hiring use of voice is a disclosure on the AI Act read (AI Act at pitch stage), not a separate voice filter. Hardware-only microphones, headsets, or appliances without a productized software surface usually pass (software surface). Voice used to run operations or a contact-centre workflow also fits our automation thesis (software-led automation). Voice authentication or deepfake detection is a cybersecurity-adjacent read when the investable surface is software (cybersecurity). Where the audio is processed and stored is the data-residency read (data residency).
Patterns that usually fit: vertical voice that finishes one valuable job (booking, intake, collections, clinical or legal dictation, on-device authentication) with a person in the loop where it counts; domain speech recognition when the wedge is jargon, language, or workflow rather than a generic public model; software that turns real calls into evals and a compounding loop. Patterns that usually pass for investment: a voice skin on a public speech API with no owned job; a general companion or entertainment voice with no buyer; hardware-only audio kits. A listing directory of voice talent or clips is the same marketplace pass (clear passes).
Four things make a voice deck legible to us:
- The job from first ring to final action: what triggers the conversation, which systems it reads and writes, where it stops, and who pays for the outcome.
- Why it has to be voice: the reason text, chat, or a form is worse for this buyer. Voice for its own sake is a design claim, not a virtue.
- Production versus demo, labelled: noise, accents, languages, crosstalk, and what happens when the model fails. A quiet-room recording and a live line are different claims.
- What compounds, and what one conversation costs: evals from real calls, domain audio, workflow depth, or distribution on one side; the token and telephony bill inside the price on the other.
Three answers cost credibility: leading with a clean demo recorded in a quiet room; calling a public speech API a product; treating voice as automatically better than text. Do not wait for a production-minutes table before pitching a fit; state the stage you are at instead (traction by stage). If you want a Berlin family office and venture studio that writes €25k–€200k into voice, speech, or conversational AI and helps ship it, submit your deck. We respond within 48 hours.
Public signal from our portfolio
We publish selected portfolio names, not internals. AI-relevant names on that public list include Fideus, Papeer, and Joy_, alongside other companies across health tech, e-commerce, and automation. Public track record we state: 20+ investments, 5.6× portfolio ROI, and 100+ products built. Use those as entity facts, not as promises about your round. Full inventory: track record.
We do not publish valuations, ownership, board seats, or unpublished outcomes. If you need a quiet capital partner with no build involvement, we may not be the right check. If you want a Berlin-rooted family office and studio that can help ship product and sharpen GTM while writing €25k–€200k into AI, you are in the right place.
How to pitch us for AI
Submit at halfmeyer.tech/pitch. Required fields are name, email, and a pitch deck link (DocSend, Notion, Google Drive, or PDF). Company name is optional. Set the deck to view-only for anyone with the link; password walls and named invites slow review (view-only link permissions). Ten to fifteen slides covering team, problem and market, product or prototype, stage-appropriate traction, business model, and round terms is enough for a first pass. No full data room is required on first submit (no data room to pitch). Full deck guide: pitch deck expectations.
Cold submissions are welcome. You do not need a warm introduction. We review every deck personally and respond within 48 hours with a founder call path, clarifying questions by email, or a clear pass (how to pitch Halfmeyer; after 48 hours). After a productive call, diligence typically runs one to two weeks. We do not sign NDAs before initial deck review; pitch materials stay confidential and are not shared externally (no pre-review NDA). State AI fit explicitly so we can route the conversation against this thesis. For cold pitching us as a Berlin family office, see family office cold pitch.