What the Evidence Says AI Changes in the First 100 Days, and Who It Changes It For

Last Updated: 10 September 2026

Before automating anything in the first 100 days, ask what the measured evidence says AI changes at work: it lifts output most for the least experienced people and barely at all for your most skilled. That means the first question is not which tool to buy but which of your workflows is routine enough to benefit, whether you can name the person who owns it, and whether you have measured how it performs today. Deploy before you have those three answers and you will not know what, if anything, improved.

What the Field Evidence Actually Shows

The best-controlled study of generative AI in a live workplace found a real, measurable gain, but concentrated almost entirely in the least experienced workers, which tells a new owner exactly where a first deployment belongs.

The best-measured study of generative AI in a real workplace found a real but uneven effect. Erik Brynjolfsson, Danielle Li and Lindsey R. Raymond studied the staggered rollout of an AI assistant across 5,179 customer support agents at a large software firm and found that access to the tool increased issues resolved per hour by 14 per cent on average, in a study covering around three million chats. The average, however, is not the story. The gain for novice and low-skilled workers was 34 per cent, while experienced and highly skilled workers saw minimal impact. The researchers' explanation is that the tool works by spreading the practices of the best performers to everyone else, so it helps most where a person has not yet absorbed those practices themselves.

For a new owner, the operational reading is narrow and useful: AI assistance behaves less like a machine that replaces work and more like a mechanism that transfers the tacit knowledge of your best people to your newest ones. That has immediate consequences for where a first deployment belongs. If you point it at work your senior people already do well, the study predicts you will see almost nothing. If you point it at routine work done by newer or less experienced staff, the same study predicts the largest measurable gain. Either way, the prediction is checkable, which is exactly what a first-100-days decision should be.

The Barrier Is Locating the Use Case, Not the Technology

UK firms say the hardest part of adopting AI is working out where it fits: the official statistics put difficulty identifying use cases ahead of both cost and skills as the most reported barrier, so producing that answer is the actual first-100-days work.

The most common reason UK firms give for not adopting AI has nothing to do with the technology's capability. In its analysis of management practices and technology adoption, the Office for National Statistics found that the most commonly reported barrier in 2023 was difficulty identifying activities or business use cases, cited by 39 per cent of firms. Cost came second at 21 per cent, and the level of AI expertise or skills third at 16 per cent. Read those in order: firms are not mainly saying "we cannot afford it" or "we lack the skills". They are saying they cannot work out where it fits.

That finding should reframe the first-100-days agenda. The scarce resource is not the tool and not even the budget; it is a defensible answer to "which activity, owned by whom, measured how?". The ONS's own numbers underline that this is a minority problem in the other direction too: its later business survey found 41 per cent of UK businesses with 10 or more employees reported no barriers to AI adoption at all, which most plausibly means many simply have not yet tried, rather than that adoption is frictionless. The firm that spends its first hundred days producing a written answer to the use-case question has done the thing that, by the official statistics, most firms find hardest.

The People Who Gain Most Are the Least Experienced

Because AI assistance helps newest team members most, point a first deployment at routine work done by newer staff, involve their direct managers in explaining it, and treat your best performers' knowledge as the thing the tool spreads rather than the thing it replaces.

The NBER finding cuts against an instinct most new owners carry: that automation is mainly about doing more with fewer, better people. In the evidence, the biggest beneficiary of AI assistance was the agent with the least experience, whose output rose by 34 per cent, while the seasoned performer gained almost nothing. For a newly acquired business, that inverts the staffing question for a first deployment. The relevant people to involve are the newest members of the team doing the routine work, and the risk to manage is what happens to the knowledge of the best performers if the tool becomes the default source of guidance.

Advisers who work on post-acquisition integrations describe the same sensitivity from the people side. McKinsey's guidance on managing talent through M&A transactions, qualitative and drawn from its advisory practice, notes that headhunters contact key staff at target and acquiring companies on the day a deal is announced, and that frontline managers, not senior leadership, are the channel employees actually trust during integration. BCG's advisory work on post-merger integration makes the parallel structural point that deferring people decisions defers the value itself. Neither source supplies a number, and none is needed: the qualitative point is that in the weeks after a close, the workforce is already unsettled, and a technology change layered on top of it lands on people who are deciding whether to stay. A first AI deployment aimed at the least experienced staff, with their managers carrying the explanation, is the version of that change most likely to be received as investment in people rather than replacement of them.

The First-100-Days Move: Baseline Before You Automate

Run a two-week baseline on one candidate workflow before anything is deployed: measure issues completed per hour, record the experience mix of the people doing the work, name an owner, and change nothing else when you deploy, so the before-and-after comparison means something.

The evidence above only becomes usable if you know what "better" means in your business, which is why the sequencing matters more than the tool choice. The practical move is a two-week baseline on one candidate workflow before anything is deployed. Worked through:

1. Pick one routine, high-volume workflow where newer staff do most of the work, for example first-response drafting for inbound customer queries. 2. Measure issues or tasks completed per hour for the people doing it, over two ordinary weeks, and note the experience mix of that group. 3. Name one owner for the workflow, and one owner for the decision about whether to automate it. These may be the same person; both must be named. 4. If you deploy, redeploy nothing else, change nothing else about the workflow, and re-measure the same measure for the same period. 5. Write the comparison down, including what did not improve.

The ONS adoption data explains why firms so rarely get to step 1: with 39 per cent citing difficulty identifying use cases, the jump from "we should use AI" to "this workflow, this team, this measure" is the actual bottleneck. And the NBER heterogeneity result explains what the baseline should capture beyond hours: if the gain concentrates in less experienced staff, a baseline that records only the team average can hide exactly the effect the study says to look for. ONS's own follow-up evidence points the same way on the skills side: among businesses citing a lack of AI expertise as a barrier, around 62 per cent respond by training or retraining existing staff, against roughly 26 per cent of those reporting no barriers, suggesting firms that take the constraint seriously treat it as a training decision, not a hiring one.

What to Produce, and Who Owns It

The output of all this is a single page: an automation candidate register. One row per candidate workflow, with the workflow name, the named owner, the pre-deployment baseline, the experience mix of the people doing the work, and, where a deployment happened, the measured comparison afterwards. The register's owner is you, the new owner, because it is the artefact that records what you decided and why; the workflow owners named inside it are the people accountable for the day-to-day measure.

The shape of a row, illustrative rather than drawn from any client engagement:

This is the same discipline Synergised applies to its own agent work: no workflow is handed to an AI system before it has a named owner, a written baseline and a defined exception path, because those three things are what make the result checkable rather than asserted.

The register is worth more than the sum of its rows. It converts the hardest step in AI adoption, the one the ONS finds most firms cannot make, into a document that already exists the next time the question comes up. It records what the evidence predicts: a gain concentrated in less experienced staff on routine work, and little movement for your experts, so a result that departs from that pattern is a finding about your workflow, not a failure of the technology. And it leaves you with something checkable rather than asserted, which is the standard every subsequent operating decision in the business should be held to.

Sources

• Brynjolfsson, Li & Raymond, "Generative AI at Work", NBER Working Paper 31161: https://www.nber.org/papers/w31161
• ONS, "Management practices and the adoption of technology and artificial intelligence in UK firms: 2023" (24 March 2025): https://www.ons.gov.uk/economy/economicoutputandproductivity/productivitymeasures/articles/managementpracticesandtheadoptionoftechnologyandartificialintelligenceinukfirms2023/2025-03-24
• ONS, "Artificial intelligence in UK businesses: 2023 to 2026" (20 July 2026): https://www.ons.gov.uk/businessindustryandtrade/business/businessservices/articles/artificialintelligenceinukbusinesses/2023to2026/pdf

Background reading (qualitative guidance, no figures):

• McKinsey, "Retain, integrate, thrive: A strategy for managing talent during M&A transactions" (19 February 2025): https://www.mckinsey.com/capabilities/m-and-a/our-insights/retain-integrate-thrive-a-strategy-for-managing-talent-during-m-and-a-transactions
• BCG, "Unlocking the Value of Talent in PMI": https://www.bcg.com/capabilities/postmerger-integration/unlocking-the-value-talent-pmi