Someone has to maintain the model. Decide who.
Prebuilt, custom-trained or configured in a console: the eight dimensions that decide it, why variability rather than volume comes first, why labelling is the real cost line, and the regression set you need on either path.
17 minYour exception queue is the real product
Straight-through rate describes the cheap half of the work. Classifying exceptions by who can resolve them, the loop that closes one, why the explanation must be derived from the failed check, and the authority an agent needs before it contacts anyone.
15 minCatch the bad scan while it's still fixable
Most pipelines validate documents after the fact: the file is stored, queued, and somewhere downstream a reviewer bounces it — blurred, wrong type, wrong holder — long after the person who could have retaken it in five seconds is gone. Validation belongs at the moment of capture, the one moment the person, the document and the software are all present and a defect costs seconds to fix instead of days. Three checks belong there, ordered deliberately: is the image usable (the gate the others depend on), is it the right document, does it belong to the right person. This piece covers each check and the specific prompt it fires, four principles that keep the gate helpful rather than hostile — run it at the capture surface, make the re-capture prompt actionable, set thresholds by risk, keep a human path for the edge — the four ways it goes wrong, and the trade-off named out loud.
12 minYour newest hire isn't human
The "digital employee" became literal in 2026 — a persistent agent that owns a process end to end, holds context across steps, calls the systems a person would call, and finishes work under bounded autonomy as a member of a team. That is a workforce question, not a tooling one, and it is distinct from two disciplines it gets confused with: agent sprawl governs dormant privilege, agent FinOps attributes cost, and neither runs the workforce. This piece is the third lens: provisioning identity and least-privilege access on day one, giving each digital worker a scope, an owner and a review date, supervising the seam with human-in-the-loop as a real layer, and offboarding with the same discipline you onboard — plus the system of record that tracks humans and agents together, the metric that retires headcount for output capacity and cost per outcome, the four ways it goes wrong, and the trade-off named out loud.
13 minYour AI vendor rents its compute
A growing share of AI inference no longer runs on the incumbent clouds — it runs on GPU-native providers, the neoclouds, that grew explosively on debt raised against their own accelerators, with revenue concentrated in a few anchor tenants and operational maturity lagging their capitalization. Your SaaS vendor's continuity now depends on a provider layer that is essential and more financially fragile than the vendor sitting on top of it. Distinct from grid-capacity supply and from cost per decision, this is provider solvency: three infrastructure risks all called "the cloud", why a provider's balance sheet reaches your SLA, the frame that diligences compute like a critical subprocessor — map the providers, check durability, make portability a contract clause, reserve capacity, read concentration as the risk it is — the four ways buyers get surprised, and the cost-versus-continuity trade-off named out loud.
13 minThe ceiling isn't price — it's the grid
The supply risk that reached the board in 2026 isn't the price of a token — it's whether there's power to produce it. Interconnection queues, transmission capacity and transformer lead times now set the maximum size of an AI data centre, and the largest operators spent the year locking continuous power under long-term nuclear contracts, small modular reactors included, because firm electricity turned scarce. This is energy supply, not the cost per decision of the piece next to it: your SaaS vendor's inference depends on compute capacity, which depends on energy that is physically capped — a chain that surfaces as a quota, a throttle or a slipped roadmap, not as a power memo. The frame that treats power as a supply input — map where inference runs, place workload by availability, reserve capacity in the contract, push the sensitive and steady to the edge, read the PPA as a resilience signal. The four ways buyers get surprised, and the capacity-versus-lock-in trade-off named out loud.
13 minThe power bill behind the token
The AI cost that bit in 2026 wasn't the token price — it was the power drawn to serve it. Inference overtook training as the dominant energy line in a model's lifecycle, data-centre power demand grew double digits a year, and electricity prices rose where compute concentrated, putting power per query inside the unit economics of every high-volume document pipeline. The efficiency frontier moved the opposite way — inference-optimised silicon, mixture-of-experts, KV-cache reuse and liquid cooling — so energy per decision became something you engineer, not accept. The FinOps-of-energy frame: energy per decision reported next to cost per decision, placement scored on intensity, SLA and jurisdiction together, and the router that right-sizes the model per call. The four ways teams still get it wrong, the efficiency-vs-latency-vs-lock-in trade-off stated honestly, and why sustainability disclosure walked out of the annual report and into the RFP as a scored line.
13 minThe seat stopped measuring the work
Software priced itself by the seat for twenty years because headcount tracked value — until the agent broke the proxy. When one operator does the work of five, seat count stops rising with the value delivered, and the agent itself never takes a seat: it runs as a service and shows up on the license report as one seat or none. The early-2026 software selloff repriced the sector as investors worked out that seats and work had come apart. What replaced per-seat isn't pure pay-per-outcome but a hybrid — a predictable base plus a variable component tied to the result — which for a document-AI vendor changes the billable unit from users and pages to decisions, cases and processed documents. The emerging frame, the four ways outcome pricing goes wrong, the alignment-vs-predictability trade-off stated honestly, and why charging for the result without knowing what it costs trades predictable revenue for variable loss.
13 minThe data changed. Nobody told the agent
Data quality is the most-cited blocker to agentic AI in 2026 — around 42% of organisations name it the number-one obstacle — but most teams fight the symptom with quality dashboards and cleanup pipelines instead of the root cause: there's no formal contract between the team that produces a dataset and the agent that consumes it. A column gets renamed, a field arrives late, an enum gains a value, and the agent decides on the new shape with logic built for the old one. The layer that fixes it is the data contract — a strictly versioned schema with freshness and completeness SLAs, a breaking-change policy, and a defined fallback on violation. The frame component by component, the four ways contracts go wrong, the velocity-vs-safety trade-off stated honestly, and why the team that skips this layer ships an agent whose reliability is hostage to every upstream sprint it never sees.
13 minCheap per run. Expensive per result
The unit economics question the board started asking in 2026 isn't what the agent costs per token or per run — it's what a completed outcome costs, adjusted for how often the agent actually completes it. An agent that's cheap per execution and fails 30% of the time has worse unit economics than one that costs 4x more and succeeds 95% of the time, because every failure carries human rework, retries and eroded trust the inference bill never shows. The frame that replaced the token report: cost-per-successful-task with success — not execution — in the denominator, a success-rate multiplier that prices rework and escalation into the unit cost, agent FinOps that sums inference plus tools plus review plus the cost of error, and an honest human baseline per task — plus the four ways the token metric sinks the board conversation, the cheaper-per-run vs cheaper-per-result trade-off stated honestly, and why "we spent X on tokens" is the sentence that loses the meeting.
13 minYou didn't buy an agent. You bought the demo
The number that repositioned agentic AI in 2026 isn't about model capability — it's about production reliability: fewer than 15% of organisations have scaled an agent to broad use, and the rest sit in pilot purgatory with agents that shine in the demo and fail on dirty input, uncooperative users and off-script scenarios. The bottleneck has almost nothing to do with model quality and everything to do with the gap between capability — the happy path — and reliability — the real case, repeatedly, under exception. The emerging discipline that closes it: success-rate SLOs per intent measured in production, continuous evaluation plus action observability plus runtime guardrails as one stack, a maturity model that gates autonomy on the numbers, deterministic fallback and human review designed in rather than patched on — plus the four ways the gap swallows a programme, the honest launch-speed trade-off, and why "which model is best" stopped being the question that matters.
13 minYour customers are training someone's moat
The moat story of 2024–2025 — "we fine-tuned our own model" — quietly expired in 2026, when frontier inference prices collapsed and open-weight models reached parity on domain extraction tasks. What stayed defensible is the data flywheel: the pipeline that turns every correction a customer makes in production — a wrong field, a reverted decision, a reclassified exception — into a retrained validator, an adjusted threshold, a refined prompt or an expanded schema. Mature flywheels report human review dropping 40–60% within 12 months without changing the base model, and a competitor entering the same account 6 months later starts behind on a curve it cannot buy. The five-component anatomy — correction telemetry, governed ingest, targeted updates, production A/B sandbox, auto-resolution rate as the KPI — the four ways the flywheel breaks, the data-rights trade-off, and why "which model do you use" is a commodity question while "how fast does your system learn my domain" decides the deal.
13 minYour vendor's exit is your problem now
In roughly twelve months document AI went from a crowded field to an active M&A board: RPA majors absorbed document-compliance vendors that had raised hundreds of millions, European ECM groups bought tax-focused OCR houses, data platforms acquired agentic capabilities — while mid-stage funding concentrated in vertical specialists. The thesis is blunt: by Q4 2026, a horizontal IDP vendor under $50M ARR with no vertical anchor and no platform partnership is a cut candidate. For the buyer that lands in three places at once — vendor-viability due diligence became strategy, the continuity clause became contract baseline, and the vertical platform with a real agentic engine earned a premium over the horizontal toolkit. The consolidator–target–survivor read on every profile, the four ways buyers and founders misread the wave, the honest trade-off, and why "what is your plan if you are acquired" now belongs in the RFP one line under "what is your SLA".
13 minSixty tools, half unused. Five are about to go
The market thesis of 2026 put a nine-figure margin pool inside one move — converting human labour into agentic software — and the side effect is vendor consolidation. The average enterprise runs 60 to 80 active SaaS tools with under half the licenses used, and once a category becomes an agent that decides and acts rather than a feature that shows a result, the buyer stops wanting five point tools and starts wanting one platform that extracts, decides, writes back and orchestrates. The hard implication: the vendor of one isolated feature is the obvious cut at the next renewal, while the platform that closes the loop captures the budget of three or four tools at once. Why the point tool goes first, the platform-fit gap that sinks the pretender — a pretty UI over a weak decision engine — the frame for what survives consolidation, the four ways a vendor loses the renewal it thought was safe, the honest trade-off, and why the buyer's RFP stopped asking what feature you have and started asking what workflow you close end to end.
13 minYour buyer stopped searching. They asked the model
The quiet demand-gen turn of 2026: the B2B buyer starts vendor research inside an answer engine, not a search box. Close to half of software buyers begin in an AI assistant, the visitors those answers refer convert well above organic, and traditional search volume is projected to fall by double digits. Generative engine optimization is the discipline of structuring content and data so the model cites and recommends you in its answer — and because no platform sells paid placement inside a generated answer, the citation is earned by citable fact, retrievable structure and off-site trust. The break from SEO is hard: instead of ten ranked positions you compete to be one of the two to seven sources the model names in a single answer. For a document AI vendor it's a funnel one — if the buyer's agent doesn't cite you, you're not on the shortlist and the demo never happens. The frame for being citable, the four ways teams get GEO wrong, the honest trade-off, and why a team that still optimises only to rank loses the top of the funnel to the one worth quoting.
13 minStop charging for the page. Charge for the decision
The quiet commercial turn in document AI in 2026 follows the turn in B2B SaaS as a whole: seat pricing fell from around a fifth of the market to about a seventh in twelve months, and pricing moved toward consumption and outcome because the agent broke the per-seat equation — one human plus three agents doesn't consume three seats, it consumes work. In document AI the per-page price repeats the same error: it charges for the input, the volume processed, when the mature buyer only wants to pay for the result — a decision approved with no rework, an invoice reconciled, a supplier cleared, a claim settled. How to define a billable outcome unit, how to keep vendor and customer incentives aligned so more automation means more revenue not less, the price floor and collar that protect margin when the hard-document mix rises, the per-outcome telemetry that makes the bill auditable, the four ways outcome pricing goes wrong, the honest trade-off between recurring-revenue predictability and value alignment, and why a vendor still billing per page loses the economic buyer who stopped buying throughput and started buying a measurable result per workflow.
13 minGrowth bought the round. Margin keeps it
B2B SaaS spent 2026 trading growth-at-any-cost for margin discipline. The median software business now grows in the low teens at a low-twenties operating margin — figures that read as weakness in 2021 and as health once cheap capital stopped subsidising unprofitable growth. Capital didn't leave the category; it moved toward retention, efficient acquisition and a gross margin the business can hold, and the vendor still pitching feature breadth loses the deal to the one that can show its unit economics. The metrics a cap-table buyer now checks before the term sheet — NDR above 110%, payback under 18 months, magic number above 1, customer concentration under 20% — how the discipline rewrites hiring toward solution and customer engineering, the four ways the growth-first reflex still sinks the model, the honest trade-off between speed and durability, and why software anchored to paid specialist work is the most defensible thesis left when money is no longer free.
13 minIt's software. The margin disagrees
The number that reframed the 2026 board conversation about AI wasn't growth or retention — it was gross margin. Agentic software carries a token tax classic SaaS never had: inference lands directly in COGS and pulls gross margin down to the 50–60% band against the 75–85% of traditional SaaS, and reasoning models inside agentic loops multiply the consumption at every level of autonomy. The implication is a valuation one, not just an operating one — a revenue multiple calibrated for SaaS margin, applied to a business with an agent's COGS, overprices the asset. Why inference is first-order COGS, where the token tax actually comes from, the levers that recover margin in points rather than vibes, why "margin per decision" belongs next to "cost per decision", the four ways the tax sinks a business, the honest trade-off between autonomy and the margin floor, and why "it's software, so 80% margin" loses credibility on the first results cycle.
13 minThe agent is ready. The data isn't
The most-repeated readiness number of 2026 keeps landing on the same verdict: only about 15% of enterprises can sustain agentic AI in production — yet 41% already run agents and roughly 60% have spent millions getting there. Read together, the figures say most of the money is sitting on a foundation that can't hold it. The barriers that dominate aren't the model: data quality and lineage (42%), compliance and sovereignty (39%), security and privacy (39%). The five layers of a foundation an agent can act on — always-on pipelines, end-to-end lineage, data contracts, interoperable governance, per-field provenance — the four ways the foundation quietly fails, the honest trade-off between waiting for a clean foundation and shipping on the one you have, and why a board grading maturity by agent count is reading the wrong number.
13 minA real agent, or RPA in a new coat
A forecast reorganised the year: more than 40% of agentic projects will be cancelled by the end of 2027 — for runaway cost, diffuse value, or inadequate risk control — and the aggravating factor has a name, "agent washing", rebranding an assistant, a chatbot or an RPA bot as an "agent" with no real autonomy. For the board, what separates the project that survives from the write-off isn't the demo — it's measurable autonomy, ROI attributed per flow, and risk control proportional to blast radius. The autonomy test that breaks the happy path on purpose, ROI per flow as the continuity gate, controls sized to what the agent can touch, the total cost that includes the reasoning loop, the four ways a board ends up in the 40%, and why the document desk is the cleanest place to run the test.
13 minBuild the moat, partner for the plumbing
One figure from the last year cut through a lot of internal pride: programs run with a specialised partner reach production at roughly twice the rate of those built purely in-house — about two thirds versus one third. The naïve reading is "outsource it", and that's wrong; so is "we build everything". The build-vs-partner call decided layer by layer — rent the model, partner for orchestration and evaluation, own the domain data and the corrected flow — the portability clause that keeps a head-start from becoming lock-in, who should own the customer's correction, the four ways a partner program quietly fails, the honest trade-off between speed now and margin later, and why the board metric that predicts who ships isn't ML headcount.
13 minThe gap isn't the model. It's the instrumentation
Roughly 95% of GenAI pilots deliver no measurable P&L impact, almost everyone has a pilot and almost nobody has a fleet in production. The cause isn't model quality and it isn't regulation — it's the three layers underneath the technology that never got built: measurement that proves the task works, integration that wires it into the flow that creates value, and ownership that keeps the system learning after go-live. The three layers in detail, the build-vs-partner signal that roughly doubles the production rate, the four ways the evaluation layer quietly rots, the honest trade-off of instrumenting before you see a number, and why "which model" is the wrong opening question.
13 minTemplates or zero-shot, decided by the tail
Two honest ways to get fields out of a document, and most of the noise comes from pretending one has to lose. Template-based extraction is deterministic, fast and near-free per page — and breaks the instant a layout drifts. Zero-shot absorbs the long tail of formats you'll never template, at a per-call cost and with quiet, plausible failures. The eight dimensions a buyer actually weighs, where each genuinely wins, the hybrid that ships — route by class and confidence, zero-shot as cold-start and fallback, templates as a cache of what zero-shot learned — and the failure mode that bites each when nobody's watching.
12 minUpskilling at the speed of automation
In 12 months agentic-AI upskilling went from an optional internal program to a board investment — consultancies certifying tens of thousands of professionals in a single deal, banks certifying whole operations functions. The dual thesis (operating agentic AI needs a different skill stack; the firm that certifies at scale first captures the talent supply and builds a switching cost), the certification-as-moat case and exactly where it rots into paper, a competence matrix by role that survives an audit, the four ways an upskilling program quietly fails, and why the document desk is the cleanest place to make the certificate predict the work.
13 minIntelligence per worker, measured honestly
Frontier firms run ~3.5x more intelligence per worker than the median, and in 2026 the metric jumped from HR slide to board agenda — arriving in the same news cycle as the AI-restructuring layoffs. It is a real measurement and an easy euphemism, one keystroke apart. The four components that make it auditable instead of anecdotal — output per FTE-equivalent, delegable-task automation rate, time recovered with a destination, quality with vs without the agent — why measuring only the numerator books six months of gain against a year of cost, and where the document desk measures it honestly first.
13 minWhere the saved hour actually goes: the productivity capture gap
The 2026 research is settled at the desk — an individual saves 5 to 9 hours a week with AI — and unsettled everywhere above it: the hour is real on the timesheet and missing on the P&L. The four exits the saved hour takes before it reaches margin — recovered slack, low-value backfill, a downstream queue that stalls, and a premature headcount cut — why measuring the individual guarantees the leak, the capture architecture that plugs each one, and where the document desk fits.
12 minFrom executor to orchestrator, by design
The real AI gain of 2026 didn't come from a better model — it came from redesigning who does what. The worker moved from task executor to orchestrator of three to five parallel agents, validating output and calibrating thresholds. Why deploy time fell ~70% and underwriting went from 10 weeks to 10 days only after the operating model changed, the new unit of work, a persona-by-persona view of the redesign, and what it means for the document layer.
13 minThe Chief AI Officer the board now audits
In 12 months the CAIO went from trend to standard — surveys put one inside roughly three-quarters of large organisations, up from a quarter a year earlier. The four questions the board now asks every quarter (model utilisation, agent SLO, regulatory exposure, unit cost), where the CAIO line sits against the CTO and the Chief Data Officer, the governance playbook that separates a real mandate from a rebranded innovation title, and what all of it means for the document layer the CAIO is on the hook for.
13 minWhen the deal is the whole network: multi-tenant network deployments
In 2026 the unit of an IDP sale stopped being one company and became a whole network — an anchor that distributes document AI to its 30,000 professionals, its correspondent banks, its brokers, its 80,000 SMB sellers. The per-node isolation architecture, the opt-in cross-node data governance that becomes the differentiator, the anchor revenue split, the four anti-patterns that leak the network and the concentration trade-off the model invites.
13 minFrontier inference at $1.50/$9: the IDP unit-cost reset
Frontier-grade inference reached $1.50 per million input tokens and $9 per million output in May 2026 — and IDP cost per document can fall from $0.18 to $0.04–0.06 without touching quality. The four pricing levers, the router that reacts in hours not sprints, the price-pass-through clause that survives the next cut, and the four anti-patterns that decouple a vendor's unit margin from the market in 90 days.
14 minProvider risk moves to the cap table: AI due diligence in 2026
AI labs hit trillion-dollar scale, the first operationally profitable quarter and IPO prep in 2026 — and enterprise procurement turned provider selection into a cap-table line. The four contract clauses every IT shop is now writing, the eight-row provider risk register, three architectural moves that turn provider risk into a hedge, and the 90-day buyer plan.
13 minCapability contracts for the post-app era
The 2026 device-and-OS shift moved the buyer one layer up: the assistant in the operating system, the notebook and the phone decides which portal to call. The three consequences for SaaS in the next two quarters, the four-artefact capability contract buyers are asking for by name — capability manifest, intent registry, tool catalog with permission scope, per-agent telemetry — and the anti-patterns that quietly retire a portal.
12 minWhen inference owns your margin: compute as the 2026 moat for document AI
Record-breaking AI capital — an US$ 122B round at US$ 852B post-money, US$ 200B+ in committed compute, an OCR floor sliding to US$ 2/1,000 pages. The three decisions a CIO and a CFO have to make now, the five-line cost build-up procurement is asking vendors to show, and the two vendor profiles most likely to get squeezed in the next 12–18 months.
12 minAgent-as-user: the post-app B2B portal
A measurable share of B2B portal sessions in 2026 originate from an assistant acting for a logged-in buyer. Six surfaces an agent actually needs, four moves in the playbook that ships, and the anti-patterns that quietly disqualify a portal.
11 minThe AI ROI gap: what the 29% do differently
Enterprise AI budgets jumped 65% to US$11.6M, but only 29% report meaningful ROI and 42% have abandoned most of their initiatives. Four traits separate programs that pay back from those that quietly stalled.
10 min