AI-enabled delivery, built to be checked.
I apply AI to requirements, backlog readiness, testing, documentation and governance reporting in regulated delivery, and I build and run these workflows myself. Every output that reaches a customer has a named person accountable for it. This page sets out what the evidence says as of November 2026, how the work is organised, and where I stop.
The constraint has moved from producing work to checking it.
Adoption is close to universal. The payoff is not, and the gap sits in the part of delivery that leaders already run: specification, review, testing and sign-off.
AI now drafts code, user stories, test packs and status reports in minutes. The amount of work arriving for review has grown far faster than the number of people able to review it properly.
Where this has been measured, in software teams, the pattern is consistent. DORA’s 2025 research found that AI raises delivery throughput while lowering stability, and calls AI an amplifier of whatever is already there. A Carnegie Mellon and Stanford study of nearly 200,000 pull requests found volume up 3.1 times against a reviewer pool up 1.5 times, while the share of changes receiving any written feedback fell from 39% to 21%.
The human review that most firms cite as their safeguard is turning into a formality. In October 2026 the FCA listed tick-box human review among its areas for improvement.
The advantage now goes to delivery organisations that plan review capacity, route work by risk and measure how often reviewers change what the AI produced. Those are decisions about governance, capacity and economics, and they belong to the delivery executive.
Where things stand, November 2026.
Nine signals I track, each with its source. Sources that sell AI products or services are marked as vendor research.
Adoption is close to universal
90% of technology professionals use AI at work. In UK financial services, 86% of firms use AI and 80% report AI agents deployed.
DORA (Google Cloud), Sep 2025 (vendor research) · Bank of England & FCA, Oct 2026
Profit impact has not moved
37% of organisations attribute any EBIT impact to AI and about 6% a significant one, both unchanged on 2025, while 80% report gains in individual productivity.
McKinsey, State of AI, Aug 2026 (vendor research)
Throughput up, stability down
Teams using more AI ship more, and their delivery becomes less stable. DORA describes AI as an amplifier of existing strengths and weaknesses.
DORA (Google Cloud), Sep 2025 (vendor research)
Review is the bottleneck
Across 8.1 million pull requests, AI-generated changes were merged within 30 days 32.7% of the time, against 84.4% for unassisted work.
LinearB benchmarks, 2026 (vendor research)
Agents sit inside delivery tools
Agents in Jira became generally available in May 2026, assignable to work items with full audit logging. The ticket now drives the work, so a vague story produces vague output faster.
The delivery layer is unmeasured
I have not found an independent, measured study of AI’s effect on requirements, acceptance criteria or status reporting. Quality engineering is the exception: a 19% average gain reported, with a third seeing little.
Capgemini World Quality Report, Nov 2025 (vendor research)
Services pricing is under pressure
Billable utilisation in professional services fell to 66.4%, the lowest in 19 years of benchmarking. Suppliers report clients asking for the same work at 25 to 30% less.
Unchecked output carries a price
In June 2026 KPMG withdrew a published report on AI that contained false case studies. Each incident of this kind strengthens a client’s hand on disclosure and verification terms.
Accountability stays with people
Regulators have added no AI-specific rules for delivery work. The ECB’s position is that “AI does not dilute responsibility”, and the FCA expects AI responsibilities mapped to named senior managers.
An AI operating system for delivery.
The reference architecture I use. Tickets and specifications drive the agents, every output arrives as a draft, and verification is a layer of its own, with a named owner at the end.
Nine places AI changes the work.
Each with the part a person still owns.
Requirements discovery
Agents draft interview guides and first-pass requirements from discovery notes and earlier programmes. The analyst owns the gaps and the trade-offs, so they surface before kickoff instead of in sprint 4.
Backlog readiness
Definition-of-Ready checks run on every story before it enters a sprint. Tickets now instruct agents directly, which makes a thin story more expensive than it used to be.
Acceptance criteria quality
Drafted against negative paths, edge cases and regulatory constraints, then challenged by a named reviewer before any agent builds against them.
Delivery planning
Dependency mapping, RAID seeding and re-baseline scenarios, with the programme manager making the call.
Documentation
Runbooks, architecture notes and decision logs generated from working artefacts, and recorded as AI-assisted.
Test preparation
Test scaffolds and data scenarios generated from the acceptance criteria and reviewed by test leads. This is the one delivery discipline with large-scale data, and the reported gains are uneven.
Knowledge reuse
Search across earlier programmes, so a new engagement starts from what the organisation has already learnt.
Governance & reporting
Status reports and steering packs drafted from delivery data, with every figure traced to its source before it goes to a client.
Review design
Review capacity planned like any scarce resource: automated checks on everything, a named reviewer where the output is consequential, and a measured rate at which reviewers change what the AI drafted.
Risks and boundaries.
Where AI helps a delivery organisation, and where I do not let it act.
● Where AI earns its place
- Drafting structured artefacts: criteria, tests, documentation, reports.
- Catching gaps and inconsistencies in requirements and backlogs.
- Synthesising prior work across programmes and customers.
- Running automated checks on every change, so human attention goes to the consequential ones.
- Shortening onboarding into complex programmes.
● Where I do not let it act
- Customer-facing communication that no named owner has read.
- Final acceptance, approval or sign-off decisions.
- Client data without the contract terms, data-location controls and DLP the client’s regulator expects.
- Figures or citations in a client deliverable that nobody has traced to a source.
- Work at a volume the team cannot review properly. When reviewers approve everything, the review has stopped working.
- Agent sessions that combine private data, untrusted content and the ability to send data out, unless a person approves the step.
What applies in EU and UK financial services.
No regulator has written AI rules specifically for delivery work. Existing accountability rules carry the weight, and these are the dates that matter. Links go to the primary source.
| Date | Source | What it means for delivery |
|---|---|---|
| 17 Jan 2025 | DORA applies (EU) | Expect bank clients to apply third-party ICT terms to AI services used on their work: data location, subcontracting, audit rights and exit plans. |
| 5 Feb 2026 | Data (Use and Access) Act 2025, s.80 (UK) | People subject to a significant decision made solely by automated means must be told, and can ask for human intervention and contest it. |
| 27 Jul 2026 | Digital Omnibus on AI, Regulation (EU) 2026/1744 | The AI Act’s high-risk obligations for Annex III uses, which include credit scoring, now apply from 2 December 2027. |
| 2 Aug 2026 | AI Act, Article 50 (EU) | People must be told when they interact directly with an AI system, at the latest at first contact. |
| 19 Oct 2026 | FCA good and poor practice on AI (UK) | No new rules. Map AI responsibilities to senior managers, keep auditable records of decisions, design for more than one model provider, and watch for human review that has become a tick-box exercise. |
| 26 Oct 2026 | Bank of England & FCA AI survey (UK) | 86% of firms use AI and 80% report agents deployed, mostly as decision support. Human approval is the control almost everyone relies on. |
| 2 Dec 2027 | AI Act, Annex III high-risk obligations (EU) | Conformity work for credit scoring and the other listed uses. A firm that repurposes a bought-in tool into one of these uses takes on the provider’s obligations (Article 25). |
How AI changes Professional Services economics.
The first visible effect has been on price. Billable utilisation in professional services firms fell to 66.4% in 2025, the lowest in SPI Research’s 19 years of benchmarking, and suppliers report clients asking for the same work at 25 to 30% less.
Senior consultants spent much of their week on coordination, documentation, status reporting and reworking artefacts. Effort-based pricing paid for those hours.
AI absorbs much of that work, and under time-and-materials the saving passes straight to the client. The firms that protect margin measure their own productivity, agree with clients how gains are shared, and move senior time to architecture, escalation and scope decisions that need experience.
Stack I run myself, as an operator
I write about this every two weeks in Delivery, after AI, my LinkedIn newsletter on how delivery, customer success and professional services organisations change under AI. Past issues are in Articles.