- 01
2027 is the year enterprise AI is judged on completed outcomes rather than deployed capability.
- 02
Automation reshapes contact center work by changing the mix, not just the volume — and the mix change is what breaks existing plans.
- 03
Integration and identity, not model choice, determine which enterprises scale in 2027.
How to read this outlook
This is a planning document, not a forecast of technology capability. Every prediction below is stated so that it can be judged at the end of 2027, and each is followed by what it means for a budget, a plan or an operating decision being made now. Where we draw on evidence we say so; where the claim is a judgement based on the programmes we run, we label it as such. The frame we use throughout is that the technology question in enterprise AI has largely been settled and the remaining questions are organisational: who is accountable, what does it cost per outcome, who supervises it, and what happens when it is wrong. Enterprises that spent 2025 and 2026 proving that models work now face a different and harder problem — industrialising something that was previously demonstrated. That is the transition 2027 is about, and it is why the winners next year will not be the organisations with the most advanced technology.
Prediction 1–3: the measure, the architecture, the pace
First, cost and quality per completed outcome becomes the standard executive measure for AI, replacing adoption, usage and time-saved metrics in the management pack. Enterprises that cannot produce that number will find their programmes deprioritised in favour of ones that can, regardless of relative merit. Second, the tool layer — governed, entitlement-aware interfaces into systems of record — becomes the recognised bottleneck and the main line item in serious AI plans, while model selection becomes a routine, reversible procurement decision. Third, the pace gap widens sharply. Organisations that built reusable layers will deploy new workflows in weeks; those that treated each deployment as a project will still take quarters, and the visible difference between the two will be attributed to talent or technology when it is actually architecture. The planning implication for all three is the same: fund the reusable layers centrally, and instrument outcomes before scaling volume.
Prediction 4–6: work, workforce and the middle
Fourth, contact centre and shared-services headcount declines in most large enterprises but by less than pilot economics implied, and the composition change is more disruptive than the reduction. Average handle time rises as simple volume automates, and organisations that fail to renegotiate that target penalise their own operations for succeeding. Fifth, a new operational function — supervision of AI in production — is formally staffed in most enterprises running agentic workflows, typically drawn from the strongest existing operators, with its own shift patterns and service level. Sixth, the compressed middle becomes visible in hiring data before it becomes visible in headcount: fewer roles that route, chase and re-key, more roles that resolve exceptions, supervise systems and hold relationships. The planning implication is that retraining must start a year before the reduction, because the roles that grow require skills that cannot be hired quickly in a competitive market.
Prediction 7–8: commercial models and the vendor landscape
Seventh, outcome-based pricing moves from experiment to mainstream in outsourced operations. Price per resolved contact, per processed case and gain-share on measured automation displace pure seat pricing for steady-state work, while seats survive for surge and specialised skills. The determinant of who benefits is not negotiation but measurement: the party with the credible pre-automation baseline and the agreed definition of resolution controls the conversation, and the clause specifying who owns the automation asset at contract end decides the value at renewal. Eighth, consolidation and disruption both accelerate in the vendor landscape, and at least one widely deployed model, framework or platform choice will be retired, repriced or made non-viable for a meaningful number of enterprises within the horizon. That is not a prediction about any particular vendor; it is a base rate. Plans should carry a migration reserve and prove portability quarterly rather than assuming stability.
Prediction 9–10: governance and the accountability reckoning
Ninth, governance moves from committee to pipeline. The enterprises that scale will encode their controls as automated release-gate checks — named owner, passing evaluation suite, authority limits configured, traces emitted, reversibility class declared — and their delivery will accelerate as a result, while committee-based governance continues to generate queues and workarounds. Tenth, a visible agentic incident at a major enterprise resets board-level expectations during 2027, and the response will not be about model accuracy. It will be about who was accountable, what limits were in place and whether the organisation could reconstruct what happened. Enterprises with an AI system inventory, named owners, authority limits in code and queryable traces will treat that moment as an operational event. Those without will suspend programmes and lose a year. The cost of being in the first group is a quarter of unglamorous work done in advance.
What we got wrong, and what would change our view
Two of our 2026 positions need revising. We expected multimodal voice quality to be the pacing item in contact centre automation; in practice integration and entitlement propagation delayed far more programmes than speech quality did, and voice capability arrived ahead of most enterprises' ability to use it. We also expected the pilot-to-production ratio to improve faster than it has, and the evidence continues to point to organisational rather than technical causes. Three things would change the view set out here. A step change in the reliability of long-horizon autonomous execution would pull the workforce effects forward by a year or more. A restrictive regulatory turn in a major market on automated decisions affecting consumers would slow customer-facing deployment and push investment inward toward employee-facing work. And a sustained rise in inference prices — which we do not expect but cannot rule out — would make unit economics decisive far earlier than the trajectory above assumes.
The 2027 planning checklist
Nine commitments, each of which can be completed inside a quarter and each of which is defensible on its own merits regardless of which predictions hold. Publish an inventory of every AI system in production with its owner and reversibility class. Fund the tool layer and agent identity as central infrastructure. Stand up evaluation as shared infrastructure and make passing it a deployment condition. Take two processes to full outcome accountability, including supervision staffing and published unit economics. Replace deflection and time-saved metrics with completion, resolution and cost per outcome. Name a human owner for every automated action. Move governance into the release path. Start the retraining programme for the compressed middle. Hold a migration reserve and exercise portability quarterly. Enterprises that complete this list will not need to be right about the future to be prepared for it — which is the only useful test of an outlook.
- 2027 is the year enterprise AI is judged on completed outcomes rather than deployed capability.
- Automation reshapes contact center work by changing the mix, not just the volume — and the mix change is what breaks existing plans.
- Integration and identity, not model choice, determine which enterprises scale in 2027.
- Outcome-based commercial models move from pilot to mainstream in outsourced operations, and measurement definitions decide who wins.
- Governance shifts from review committees to automated release gates, and that shift accelerates delivery rather than slowing it.
Questions leaders ask us
- How is this outlook different from a technology forecast?
- It is a planning document. Every prediction is written to be judged at the end of 2027 and is paired with a decision a leader is making now — a budget line, a metric change or an operating commitment.
- What is the single biggest planning error going into 2027?
- Planning for volume change without planning for mix change. Automation removes the easy work first, which makes the remaining work harder, longer and more skilled — and breaks forecasts built on averages.
- If we can only do three things next year, what should they be?
- Build the governed tool layer, stand up evaluation as shared infrastructure, and take one process to full outcome accountability with published unit economics. Everything else is easier once those exist.
- Is model choice really not strategic?
- It is consequential but reversible, provided evaluation suites and tool definitions live outside the vendor. The strategic choices are the ones that are expensive to undo: integration, identity, accountability and measurement.
- How often is this outlook updated?
- Annually, with the following year's edition published in the fourth quarter, plus interim revisions when a prediction is clearly overtaken by events.
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] The EU AI Act phases obligations for general-purpose and high-risk AI systems across 2025–2027. Regulation (EU) 2024/1689 (Artificial Intelligence Act) — Official Journal of the European Union, 2024
- [3] Recognised AI risk management practice organises controls around govern, map, measure and manage functions. AI Risk Management Framework (AI RMF 1.0) — NIST, 2023