Talomnia Investor Market & Venture Opportunity Analysis

Executive summary

This research answers the investor's question, not the product one: is the market large enough, growing fast enough and reachable enough, is the position differentiated, and is there a realistic path to a large company. The published Talomnia Workforce Market Research answers a different question — what to sell first and to whom — and is preserved as its own standalone artifact.

Conclusion: speculative pre-seed. Not "interesting pre-seed", not "seed-ready". What is demonstrated is a governed execution pipeline with a machine-checked evidence trail; what is not demonstrated is anything commercial. The reasons NOT to invest are listed explicitly in the final section — a mandatory part of this document, not a footnote.

• Market size is built by two independent method families and they disagree by ~2.5x — that is a finding, not noise: provider revenue includes government purchases, household sales, exports and capitalised output, none of which appear in business intermediate purchases.
• The right denominator for Talomnia is the demand-side one: $0.44-0.49T (US), not the $1.23T supply-side figure.
• The "AI agents" category carries two mutually exclusive definitions with a ~25x gap. Neither is used here as a TAM.
• The platform scenario's central economic mechanism — reuse economics — is not demonstrated, and the first-party measurement supports neither direction: the ratio moves from 1.66x to 0.73x with the choice of population alone.
• The public comparables are being repriced downward: Endava ~0.17x revenue, Concentrix ~0.16x, Fiverr ~0.8x.

Back to contents

Market definition: the wedge and the platform

The analysis is not limited to the "AI agents" category. Talomnia intersects several markets, and they must be distinguished because they are nested inside one another rather than adjacent.

• The wedge market (what Talomnia Workforce sells today): fixed-scope research, technical and product analysis, software development, AI-first consulting. Measured by professional-services statistics.
• The potential platform market (what Professional Capability Infrastructure would create): reusable professional capabilities as infrastructure. Measured by no existing statistic and carrying no evidence today.

Structural warning: the "AI agents", "agent orchestration" and "enterprise AI spend" segments are nested, not adjacent. Summing the three triple-counts. Gartner's "agent builder platforms" sub-segment ($5.0B, 2026) sits inside its DSML segment ($29.9B), which sits inside total AI spend ($2,595,667M).

Back to contents

The 25x gap: a measurement problem, not a contradiction

The largest number circulating in this category carries two mutually exclusive definitions. The source that compiles both says so outright: "That 25x gap is not a contradiction. It is a measurement problem."

For 2025 the gap between Gartner ($86.4B) and the report-vendor consensus ($7.06-7.92B) is ~11x. These may neither be summed nor averaged. Neither is used as a TAM in this model, because Talomnia does not sell agent software — it sells completed professional work, which is measured by different instruments entirely.

Back to contents

Method A — top-down, supply side

Family: government statistics. Measures the revenue of providers of this work, from the US Census economic surveys. Formula: TAM_A(US) = Revenue(NAICS 5415) + Revenue(NAICS 5416) + Revenue(NAICS 54191).

Sensitivity: narrow (consulting and research only) $0.47T; stated $1.23T; broad (all of NAICS 54) $3.11T. The spread is driven by scope choice, not measurement error.

Global extrapolation is deliberately not performed. No government agency publishes a global professional-services total. Eurostat publishes value added (EUR 779.3B, NACE M, 2023), which is not revenue. The only global figures are aggregators, disagreeing by 5.3x ($1.20T vs $6.37T), both low confidence and both excluded under the no-single-aggregator rule. Any global figure here would be an assumption wearing a measurement's clothes.

Back to contents

Method B — bottom-up, demand side

Family: government statistics again, but a different instrument — business expenditure surveys and national input-output accounts rather than provider revenue. That is the independence the acceptance criterion requires: Method B can be wrong in ways Method A cannot, and vice versa.

Two independent demand-side instruments, built by different Census and BEA programmes five years apart, agree within 12%. Sensitivity: conservative (BEA 2017) $0.44T; stated (Census 2022) $0.49T; grown at the measured +6.5% Y/Y to 2026, $0.63T.

Back to contents

The headline disagreement — and why it is a finding

The methods disagree by ~2.5x and the cause is identifiable rather than mysterious. Provider revenue exceeds business intermediate purchases because it additionally contains: sales to government; sales to households ($114,273M against commodity 5412OP); exports ($224,890M and $51,912M); and capitalised output ($767,254M and $342,948M booked as intellectual-property investment rather than as an expensed input). The last is the largest single component of the gap.

Which number is the right denominator for Talomnia? The demand-side one. Talomnia sells B2B work billed against a customer's operating budget with acceptance criteria. It does not sell to households, is not an exporter of record, and its output is not capitalised on the client's balance sheet. An investor should anchor on ~$0.44-0.49T (US), not on $1.23T — and should treat any pitch leading with the larger number as having chosen the flattering instrument.

Back to contents

SAM and SOM — not computed, and that is the answer

SAM = TAM_B x the share of that spend that is fixed-scope and outcome-specifiable x the share of buyers who will transact with a pre-commercial vendor. Two of the three factors have no evidentiary basis at any source class: the Census SAS class-of-customer table does not cover NAICS 54. The honest output of a model with unknown factors is unknown, not a plausible product.

What can be bounded is the buyer count: 285,007 US employer firms with receipts >=$10M (Census SUSB 2022, computed — no official row publishes this cutoff), of which 28,581 are in NAICS 54 itself; 303,073 EU enterprises with 50+ employees (Eurostat SBS 2023); 44,545 UK enterprises with turnover >=GBP10m (ONS IDBR).

At the current stage SOM is bounded not by market size but by delivery capacity and by a sales cycle Talomnia has never run. Modelling it from TAM is exactly the error this document exists to avoid. Talomnia has zero external paid engagements; every input that would make SOM meaningful is a quantity only the first paid cases can produce. A SOM presented today would be an assumption stack, and an investor should treat any such number from any vendor at this stage the same way.

Back to contents

Why now

Five structural drivers, each with evidence. A sixth and most-quoted one — "95% of AI pilots fail" — is deliberately NOT used: the primary source returned 403 and was never read.

• Enterprise generative-AI spend grew from $1.7B (2023) to $11.5B (2024) to $37B (2025), a 3.2x year-over-year increase (Menlo Ventures primary research, n=495).
• Agent capability measurably improved: real-world task success went from 20% (2025) to 77.3% (2026) on the Stanford HAI benchmark.
• Large capital confirms category appetite: AI took $211B, about 50% of all global venture funding in 2025, and $242B - 80% - in Q1 2026.
• Incumbents are buying the capability rather than building it: Accenture/Faculty, AlixPartners/Artium, IBM/Hakkoda, Cognizant/3cloud - a repeated, dated pattern.
• Trust became the competitive battleground because incumbents publicly lost it: Deloitte refunded more than half of an A$440,000 fee for a report with fabricated legal references; a KPMG report had 40 of 45 citations fabricated; the Financial Times verified fabricated footnotes across four PwC reports.

The fifth driver is the direct justification for Talomnia's positioning. But it cuts both ways: the same failure raises distrust of every vendor of AI-delivered work, including new entrants with no track record.

Back to contents

Competitive landscape

The map is not limited to direct analogues. Six groups, each labelled by what it is to Talomnia: a competitor, a substitute, or a layer above.

Back to contents

Competitive matrix

Every cell without evidence reads "unknown". A cell estimated without evidence is a fabricated cell; the value of this matrix to an investor is precisely that its blanks are real.

A notable asymmetry: pricing is unpublished by almost everyone, Talomnia included. Published price lists exist for tools (Cursor, LangChain, Replit) and for one service (Superside). That is consistent with a market that has not yet agreed on the unit in which an outcome is sold.

Back to contents

Funding and venture landscape

The answer to "is there demonstrated venture appetite for this category" is yes, but it is extremely concentrated.

The concentration is both confirmation of appetite and a warning: capital is going overwhelmingly to model labs, not to the services built on them. Rounds inside the AI-executed-work category itself are real and dated: Sierra $950M at a $15.8B valuation (May 2026); Harvey $200M at $11B (March 2026); Decagon $250M at $4.5B (January 2026); Legora, OpenEvidence, Abridge, EvenUp and Crosby all with verified 2024-2026 rounds.

Multiples are almost never disclosed. Across the whole collected corpus only three revenue multiples were explicitly stated by a source: OpenEvidence ~70x, Hebbia ~54x and Invisible 8.3x. The gap between the first two and the third is the most informative quantity in the data, because Invisible is the only one of the three that is explicitly a services business.

Back to contents

Comparable companies — and why they are comparable

Comparability is explained per company. None of them is called a direct competitor.

Back to contents

Market map

Two axes explain the positions: what is sold (tool → delivered outcome) and reach (single vertical → cross-domain). Talomnia's assumed position is the top-right: delivered outcome, cross-domain.

The hypothesis "cross-domain accountable work + Knowledge Contract + reusable capabilities + execution evidence" is not treated as unique merely because no exact match was found. The likelier explanation for the empty slot is not market blindness but that cross-domain generalisation is harder and monetises worse than vertical depth: all four well-funded companies in the top-left chose depth, and the market paid for that choice.

Back to contents

Defensibility: graded, not asserted

A mechanism that exists only as a design intention is graded Hypothesis or Not demonstrated and is not called a moat. Score: 2 Current, 3 Emerging, 3 Hypothesis, 8 Not demonstrated.

Both Current mechanisms are about provable execution, not accumulated advantage. That is the honest summary of the position: Talomnia can show its work, and cannot yet show that showing its work compounds.

Back to contents

Venture scale: scenarios, not a forecast

Back to contents

Bear case: the reasons not to invest

Written at the same depth as any bull argument, because an investor who cannot see this is being sold to rather than informed.

• The productivity premise is contradicted by a randomised trial. METR (2025-07-10), 16 experienced developers, 246 issues: "when developers are allowed to use AI tools, they take 19% longer to complete issues". Worse for the category's self-reporting: "developers expected AI to speed them up by 24%, and even after experiencing the slowdown, they still believed AI had sped them up by 20%".
• The public comparables are being repriced downward, hard. Endava -77.9%, Fiverr -59.6%, Concentrix -48.3%, Globant -43.6%, Upwork -41.5%, EPAM -39.6% (observed 2026-08-19). The one exception, Innodata (+71.3%), sells training data, not AI-delivered work.
• Gross margin is not computable from the current instrument. Of 653 ledger entries, 582 state human review time as zero, 64 record it honestly as unmeasured, and only 7 carry a measured value; those 89% zeros are an absent measurement written as a number. Any margin derived from this dataset omits human cost entirely and is an upper bound, not a margin.
• Category insiders name this exact failure mode. Emergence Capital, "Mirage Product Market Fit": "fast revenue growth, strong customer retention, but the majority of the service is still being delivered by humans, not AI... You haven't built an AI-native services business. You've built a services business with the wrong kind of funding." Thresholds: below 50% gross margin is a services business misclassified; 70%+ is needed for software multiples. Bessemer's own "Supernova" cohort at ~$40M ARR in year one carries 25% gross margins - inside the misclassified band by Emergence's own test.
• Incumbents are buying the capability rather than losing to it: Accenture/Faculty (completed 2026-03-16, Faculty's CEO became Accenture's CTO), AlixPartners/Artium, IBM/Hakkoda, Cognizant/3cloud. The consolidation path may close before an independent reaches scale.
• Model providers are entering services directly. OpenAI acquired the consulting firms Tomoro and Convogo and launched a $4B OpenAI Deployment Company with Bain, Capgemini and McKinsey as integrators. The layer Talomnia occupies is being occupied from above.
• Trust is the battleground and the category keeps losing it. 11x displayed logos of companies that were not customers (ZoomInfo on record: "we did not give them permission to use our logo... we are not a customer"); Builder.ai used humans for work marketed as AI and entered insolvency. A new entrant inherits the category's credibility discount without having earned any of it.
• Zero commercial validation: no paying customer, no acceptance data, no willingness to pay, no repeat purchase, no measured sales cycle.
• The legal entity is not registered and no security certification is held, while competitors hold SOC 2 Type II, ISO 27001, ISO 42001, HIPAA and FedRAMP. Both block the enterprise procurement the model targets.
• Key-person risk: a single operator, with no contributor ecosystem and no evidenced team redundancy.

Back to contents

Bull / Base / Bear

Bull - what would have to be true. Reuse economics reverses and shows a real cost decline; the evidence trail becomes a purchasing criterion after the Deloitte, KPMG, PwC and EY defect wave; Talomnia wins and repeats fixed-scope engagements at >=70% gross margin with human review measured and included; cross-domain generalisation holds where competitors stayed vertical. Ceiling if all four hold: Harvey $11B, Sierra $15.8B - both vertical, both with disclosed ARR.

Base - the most realistic at the current evidence level. A small, credible services business that wins design-partner work on the strength of its trail, grows at services economics, and is valued on services multiples (0.2x-1.5x revenue per the public set). Venture scale is not the base case on today's evidence.

Bear - under what conditions it does not work. Reuse never compounds; margin lands below 50% once human review is counted; incumbents and model labs close the layer from both sides; enterprise procurement blocks a vendor with no certifications and no registered entity; the first paid engagement never converts to a second.

Back to contents

Exit paths — strategic logic only

No claim is made that any named company will acquire Talomnia. The strategic logic of the possible paths is what is analysed.

Back to contents

Evidence Gates — what would de-risk this, in order

Concrete future proofs. Values are not fixed where the research cannot justify them - the metrics and the ranges worth testing are proposed instead.

Order matters: EG-1 and EG-2 subsume most of the others. Without them, no figure in the scenarios section is testable.

Back to contents

Investor recommendation

Rating: speculative pre-seed. Not "interesting pre-seed", not "seed-ready".

What is already demonstrated. A governed execution pipeline producing a machine-checked evidence trail, with gates that actually refuse. The production of this very document was blocked twice by its own ledger validator and once by a projection status-honesty gate. The trail is real, unusually rigorous for this stage, and directly answers the failure mode the Big Four have publicly demonstrated.

What remains hypothesis. Everything commercial: reuse economics, margin, willingness to pay, repeat purchase, cross-domain generalisation, switching costs and every network effect.

What could kill the thesis. Reuse never compounding (the platform case); margin below 50% once human review is counted (the whole case); model labs and consulting incumbents closing the layer from both sides - both now evidenced with dated transactions.

• there is no revenue, no customer and no acceptance evidence of any kind
• the central economic mechanism is not merely unproven but unmeasured: the ratio reverses direction with the choice of population, and the instrument cannot separate reuse cost from the work beside it
• gross margin cannot be computed from the instrument the company itself uses
• the public comparables for the realistic outcome trade at 0.16x-0.8x revenue and are falling
• a randomised trial contradicts the productivity premise the category sells
• the category's credibility has been damaged by others and a new entrant inherits the discount
• the legal entity is not registered and no security certification is held, which blocks the targeted enterprise buyers
• capital in this category is concentrating in model labs, not in the services built on them

Recommended stage: pre-seed, sized to reach EG-1 and EG-2 only. Milestones before fundraising: EG-9 (entity), then EG-1 (first paid case), then EG-2 (measured margin including human review). Capital requirement is not estimated - the research cannot justify a figure without a measured sales cycle and delivery cost, and an estimate here would be exactly the assumption this document refuses everywhere else.

Back to contents

Sources and evidence limits

22 of 22 published citations were verified by this lane with a live fetch on 2026-08-20 with the HTTP status recorded. Source priority: official filings and government statistics first, then major consultancies and business media; market-report aggregators only as low-confidence evidence and never as the sole basis for a market size.

A search-tool limitation that must not be read as absence. After exhausting the search budget, two research lanes fell back to a news index that is recency-weighted: it returns "no results" for most 2024 and many 2025 events that certainly happened. Every "not found" produced through that channel means the index did not surface it, NOT that the event is absent from the record. Marked unknown for this reason rather than for absence of evidence: Sierra's outcome-based pricing figures, Upwork and Fiverr take rates, the Replit database-deletion incident, the Cursor support-bot incident, the FTC defendant list, and the Moffatt v. Air Canada damages figure. All are widely reported; none could be fetched this session; none are asserted here.

What this lane did not verify. The acquisition press releases underpinning the exit-path section were fetched with status 200 by the research lanes, but the addresses this lane reconstructed returned 404 and the search budget was exhausted. Those facts remain in the text with their source class stated, but are excluded from the published sources list, which carries only addresses this lane verified itself. A 404 on a reconstructed address is evidence that the address is wrong, never that the transaction did not happen.

No McKinsey figure appears anywhere in this document. A snippet asserting "72%" was available and was deliberately not entered as evidence, because the host could not be reached to confirm it.

Back to contents

Goal

Give an investor a factual basis to answer whether Talomnia's market is large enough, growing fast enough and reachable enough, whether the position is differentiated, and whether a realistic path to a large company and a return exists. The research must not set out to prove a predetermined positive conclusion: a negative result is admissible.

Back to contents

Methodology

Desk research on 2026-08-20, executed by 36 accounted agents across four lanes (top-down, bottom-up, funding and comparables, competitive matrix and exits). Market size is built by two independent instrument families: provider revenue from the US Census economic surveys, and business expenditure from the Census expenditure survey and the BEA input-output accounts. Every figure carries its source, class, year, confidence and sensitivity range. Market-report aggregators are used only as low-confidence evidence and never as the sole basis for a size. Before publication every published source was fetched live by this lane with its HTTP status recorded; 429 and 403 mean the server answered and the source is reachable, and only 404 and NXDOMAIN prove absence.

Back to contents

Results

The right denominator for Talomnia is the demand-side one: $0.44-0.49T (US), not the $1.23T supply-side figure; the 2.5x disagreement is explained by government purchases, household sales, exports and capitalised output. Two independent demand-side instruments converge within 12%. The AI agents category carries two mutually exclusive definitions with a 25x gap and is not used as a TAM. Venture appetite is demonstrated but extremely concentrated: OpenAI and Anthropic took 43% of H1 2026 funding. Defensibility: 2 Current, 3 Emerging, 3 Hypothesis, 8 Not demonstrated; reuse economics, the platform scenario's central mechanism, is not demonstrated: the create-to-reuse ratio moves from 1.66x to 0.73x purely with the choice of population, and the attribution method in use cannot separate reuse cost from the work co-located with it.

Back to contents

Limitations

No global TAM is built: no government agency publishes a global total, and the only global figures available are aggregators disagreeing by 5.3x. SAM and SOM are not computed: two of the three SAM factors have no evidentiary basis at any source class, and at this stage SOM is bounded by delivery capacity and a sales cycle Talomnia has never run. All own-economics data is pre-commercial self-measurement; there are no commercial cases. Gross margin is not computable: of 653 ledger entries, 582 state human review time as zero, 64 record it honestly as unmeasured, and only 7 carry a measured value — those 89% zeros are an absent measurement written as a number, not the observation "no human took part". Input vintages are mixed and not normalised. Several acquisition press releases were not re-verified by this lane and are excluded from the sources list. The news index two lanes fell back to after exhausting the search budget is recency-weighted - not found means an index limit, not absence from the record.

Back to contents

Conclusions for Talomnia

Speculative pre-seed. What is demonstrated is a governed execution pipeline with a machine-checked evidence trail and gates that actually refuse; what is not demonstrated is anything commercial. Recommended stage is a pre-seed sized to reach EG-1 (first paid case) and EG-2 (measured gross margin including human review) only. Capital requirement is not estimated: no figure can be justified without a measured sales cycle. The reasons NOT to invest are stated explicitly and include zero commercial validation, a central reuse mechanism that remains unmeasured because the instrument in use cannot isolate it, an uncomputable margin, falling comparable multiples, and a randomised trial contradicting the category's productivity premise.

Back to contents

Sources

• https://www.census.gov/services/qss/qss-current.pdf
• https://www2.census.gov/programs-surveys/sas/tables/time-series/sas-latest/sas-22.xlsx
• https://www2.census.gov/programs-surveys/susb/tables/2022/us_6digitnaics_rcptsize_2022.xlsx
• https://apps.bea.gov/industry/iTables%20Static%20Files/AllTablesSUP.zip
• https://softwarestrategiesblog.com/2026/02/26/roundup-of-agentic-ai-forecasts-and-market-estimates-2026/
• https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/
• https://emcap.com/thoughts/mirage-product-market-fit-the-silent-killer-of-ai-native-services-companies
• https://bvp.com/atlas/the-state-of-ai-2025
• https://www.sequoiacap.com/article/services-the-new-software/
• https://menlovc.com/perspective/2025-the-state-of-generative-ai-in-the-enterprise/
• https://www.cityam.com/kpmg-report-on-ai-found-riddled-with-ai-hallucinations/
• https://techcrunch.com/2025/03/24/a16z-and-benchmark-backed-11x-has-been-claiming-customers-it-doesnt-have/
• https://en.wikipedia.org/wiki/Builder.ai
• https://cognition.com/blog/windsurf
• https://stockanalysis.com/stocks/dava/
• https://stockanalysis.com/stocks/fvrr/
• https://stockanalysis.com/stocks/cnxc/
• https://stockanalysis.com/stocks/inod/
• https://ec.europa.eu/eurostat/statistics-explained/index.php?title=Structural_business_statistics_overview
• https://sierra.ai/blog/year-two-in-review
• https://www.langchain.com/blog/series-b
• https://news.crunchbase.com/venture/global-startup-exits-ipo-ma-soar-ai-q2-h1-2026/

Back to contents

How this was performed

/en/workflows/investor-market-venture-analysis

Back to contents

Versions

v0.1.0 — 2026-08-20