- 01
By 2027 the unit of enterprise automation is a completed outcome, not a completed prompt — and budgets will be reset around that measure.
- 02
The scarce asset is not models but the tool layer: governed, entitlement-aware interfaces into systems of record.
- 03
Agent supervision becomes a named operational function with staffing, shift patterns and a service level, in the way network operations did two decades ago.
The thesis
The defining shift between now and 2027 is not model capability. It is the movement of enterprise AI from a layer that helps a person do work to a layer that completes work and reports back. Assistive AI produced real but bounded value: shorter handle times, faster drafting, better search. It left the operating model untouched, because a human still owned every step. Agentic execution changes the ownership of the step itself, and once a system owns a step it inherits everything that surrounds ownership in an enterprise — entitlements, audit, error budgets, escalation paths, accountability when it goes wrong. Our view is that 2027 is the year that inheritance is settled in most large organisations, not because the technology arrives then, but because the governance, integration and workforce questions take roughly this long to work through. Enterprises that begin that work in the next four quarters will spend 2027 scaling. Enterprises that wait for the questions to be answered elsewhere will spend 2027 doing what the first group did in 2026.
What the architecture looks like when it settles
The durable pattern we see forming has four properties. First, a governed tool layer: every action an agent can take is exposed as an interface that enforces the caller's entitlements, applies authority limits in code rather than in instructions, and returns predictable errors. Second, retrieval grounded in versioned policy — the system reasons over the rule in force for that jurisdiction on that date, and cites it. Third, orchestration that treats partial failure as normal, holds state so an interrupted case resumes rather than restarts, and knows when to hand to a human. Fourth, evaluation as infrastructure, not as a phase: golden sets that run in the deployment pipeline, traces on every case, and quality reported as a trend rather than asserted in a review. Model choice sits above all of this and becomes increasingly interchangeable. That is the point most 2026 architectures get backwards. The organisations that invested in the tool and evaluation layers will be able to adopt each new model generation in weeks; the organisations that invested in a specific model will rebuild.
The operating model changes before the org chart does
Expect the sequence to be: new measures, then new roles, then new structure. The measure that reorganises everything is cost and quality per completed outcome — per resolved case, per processed claim, per fulfilled request — replacing activity measures like tickets handled or calls answered. Once that measure is in the management pack, three roles appear. An agent product owner, accountable for the outcome quality of a workflow the way a product manager is accountable for a feature. An agent operations function, watching live behaviour, triaging regressions and managing the escalation queue. And an evaluation owner, who maintains the test sets that define what good means and defends them from being quietly weakened. Only after those roles are working does structure change, usually by pulling shared services, contact centre operations and parts of IT service management under one accountability for end-to-end resolution. Leaders who attempt the structural change first — a reorganisation announced before the measures exist — reliably lose a year.
Work and workforce: the honest version
The credible 2027 outcome is not mass displacement and not business as usual. It is compression of the middle. Simple, high-volume, well-documented work moves to agents faster than most workforce plans assume. Complex, judgement-heavy and relationship-heavy work becomes more valuable, not less, and the people doing it need better tools and deeper authority. What compresses is the layer in between: the roles that mainly consisted of routing, chasing, re-keying and assembling context. Two implications follow for planning. First, hiring profiles change before headcount does — enterprises will need fewer people who process and more who resolve exceptions, supervise systems and hold customer relationships, which is a retraining problem long before it is a redundancy problem. Second, the supervision workload is real and permanent; every serious agentic deployment we see creates a smaller but skilled team whose job is to keep it honest. Plans that book savings without funding that team overstate the benefit and understate the risk.
Where the 2027 risk actually sits
Three risks dominate, and none of them is model hallucination in the way boards currently discuss it. The first is accountability drift: an agent takes an action inside a process where nobody has been named as accountable for the outcome, and the gap only becomes visible during an incident. The fix is administrative, not technical — every automated action needs a named human owner recorded before go-live. The second is silent quality decay. Agent quality degrades gradually as upstream content changes, prompts are edited and models are swapped, and without continuous evaluation the decay is discovered by customers. The third is concentration: a single vendor's model, platform or agent framework embedded so deeply across workflows that switching becomes a multi-year programme. Enterprises should assume at least one forced migration between now and 2029 and design the tool and evaluation layers so it is survivable. Regulatory exposure is a fourth pressure rather than a risk in itself, and it is largely a documentation discipline: the organisations already keeping traces, evaluation records and decision logs will find compliance a reporting exercise.
Implications by role
CIO: the 2027 constraint is integration surface, not AI budget — fund the tool layer and the identity work now, because it is the long-lead item and it is reusable across every workflow. Chief AI Officer: move the mandate from experimentation to production standards; the highest-value artefact you can own is the evaluation and release process every team must pass. COO: pick two end-to-end processes and take them all the way to outcome accountability rather than spreading pilots across ten; depth is what produces a defensible number. CX leader: assume containment stops being the headline metric and resolution quality replaces it, and start instrumenting for that now. CFO: insist that every AI business case includes run cost, supervision headcount and the cost of a forced platform migration, and require a unit-economics line rather than a productivity percentage. CHRO: begin the retraining path for the compressed middle in 2026, because the roles that grow require skills that take a year to build.
What to do in the next four quarters
A practical run-up to 2027 has four moves, one per quarter, and each is deliberately small enough to complete. Build or formalise the governed tool layer for the two or three systems of record that appear in most of your workflows, with entitlement propagation and authority limits enforced in code. Stand up evaluation as shared infrastructure so every team ships against the same bar. Take one process to full outcome accountability, including the supervision staffing and the run cost, and publish the unit economics internally so the organisation learns what a real number looks like. Then set the portfolio rule for 2027: no new agentic workflow enters production without a named human owner, an evaluation suite and a stated cost per outcome. None of this depends on which model leads next year, which is precisely the point. The enterprises that look prepared in 2027 will be the ones that made these four unglamorous moves while everyone else was comparing benchmarks.
- By 2027 the unit of enterprise automation is a completed outcome, not a completed prompt — and budgets will be reset around that measure.
- The scarce asset is not models but the tool layer: governed, entitlement-aware interfaces into systems of record.
- Agent supervision becomes a named operational function with staffing, shift patterns and a service level, in the way network operations did two decades ago.
- Organisations that keep a human decision point on every irreversible action will move faster, not slower, because approval friction falls away everywhere else.
- The competitive gap will open between enterprises that industrialised evaluation and those that still judge AI quality by demonstration.
Questions leaders ask us
- Is 2027 realistic, or is this another pilot cycle?
- The technical capability already exists in 2026; what has not settled is governance, integration and accountability. Those take three to six quarters in a large enterprise, which is what puts credible scale in 2027 rather than sooner.
- Do we need to pick an agent framework now?
- No. Framework choice is the most reversible decision in the stack. The tool layer, identity model and evaluation suite are the durable investments, and they make a later framework change survivable.
- How do we budget for something this uncertain?
- Fund the reusable layers as infrastructure with a fixed annual envelope, and fund workflows individually against a stated cost and quality per completed outcome. That separates the bet you can defend from the bets you are testing.
- What is the single most common mistake going into 2027?
- Booking savings without funding supervision. Every production agentic workflow needs a small skilled team watching it, and business cases that omit that team overstate the benefit and hide the risk.
Sources
- [1] The large majority of enterprise generative AI pilots never produce a measurable production outcome. The GenAI Divide: State of AI in Business — MIT NANDA / Project NANDA, 2025
- [2] Recognised AI risk management practice organises controls around govern, map, measure and manage functions. AI Risk Management Framework (AI RMF 1.0) — NIST, 2023