Agents
The AI agent gap is widening. What the August 2026 data says SMEs should do about it
McKinsey, Salesforce and Google all shipped news in late August 2026 that changes the case for AI agents in a 10 to 250 person business. Here is what it means.
The short answer
Three findings published in late August 2026 point the same way. Large organisations scaling AI agents rose to 40 percent while smaller organisations stayed flat at 22 percent (McKinsey, 25 August 2026). Professional and business services firms were among the slowest to adopt agents but reached return fastest, at around 6.5 months, on the back of clean data, narrow scope and clear human escalation (Salesforce, 27 August 2026). And Google Cloud introduced pay as you go pricing and cost controls for Gemini Enterprise and agents on 26 August 2026, which removes the licence commitment that stopped many smaller firms piloting at all. The correct response for an SME is not to move faster. It is to move narrower.
Key takeaways
- The adoption gap is between large and small firms, not between fast and slow ones.
- Slow, narrow starters reported the fastest payback. Breadth is what kills agent programmes.
- Usage based pricing has removed the main financial reason not to run a first pilot.
- EU AI Act enforcement is live now, but high risk obligations are deferred to December 2027.
- UK firms have a new £100m public funding competition worth watching, announced 31 August 2026.
Four things landed in the last week of August 2026 that, taken together, change the calculation for a business of 10 to 250 people deciding whether to build its first AI agent. None of them is a model launch. All of them are about economics, discipline and rules.
1. The gap is size, not speed
McKinsey's State of AI, published 25 August 2026, found that 40 percent of organisations with revenue above one billion dollars are now scaling AI agents, up from 27 percent a year earlier. The share of smaller organisations doing the same was flat at 22 percent.
That is a widening gap, and it is worth being precise about why. It is not that smaller firms are slower to try things. It is that scaling an agent requires something most SMEs have not funded: an owner, an evaluation set, monitoring and a maintenance budget. Large firms are not better at prompting. They are better resourced at the boring part.
2. Slow starters got the fastest return
Salesforce published an agentic AI leaders survey on 27 August 2026 with a genuinely awkward finding for the move fast crowd. Professional and business services firms were among the slowest sectors to deploy agents, and among the fastest to see meaningful return, at roughly 6.5 months. The reasons given were clean data, narrow agent scope and clear human escalation paths.
This matches what we see in the field. The businesses that stall are the ones that green light five agents at once across five departments. Attention splits, no single workflow gets a proper test set, and nothing reaches the quality bar. The businesses that succeed pick one high volume workflow, define what must never happen, and ship it.
| Approach | Typical first 6 months | Why |
|---|---|---|
| Five pilots at once | Nothing in production | No workflow gets a real evaluation set or an owner |
| One workflow, evaluated | One agent live, measurable | Scope is small enough to test properly and fix quickly |
| Buy licences, train nobody | Low usage, no result | Capability, not access, is the constraint |
3. The pricing barrier just dropped
On 26 August 2026, Google Cloud introduced flexible pay as you go billing plus cost visibility and control tooling for Gemini Enterprise and agents. The stated driver was enterprise anxiety about unpredictable agent spend, but the practical beneficiary is the smaller firm that could never justify a seat based annual commitment for a pilot.
Usage based pricing cuts both ways. It removes the upfront barrier and introduces a variable one. Agree a monthly ceiling before launch, instrument token and call volume from the first day, and treat a cost spike as a bug rather than a surprise on the invoice.
4. The rules, honestly stated
The European Commission's AI Office now holds live enforcement powers, with penalties reaching 3 percent of global turnover, and coverage on 28 August 2026 noted that the regime is real even though no cases have yet been brought. Meanwhile the AI Omnibus regulation that entered into force on 27 July 2026 deferred high risk obligations to December 2027 and extended lighter touch treatment to companies up to 750 employees and 150 million euro turnover.
For a UK business selling into the EU, the honest reading is this. Transparency duties matter now. High risk compliance is a 2027 problem, and many growth stage firms that assumed they were out of scope for relief now qualify for it. Do the exposure check. Do not build a compliance function for it yet.
5. UK funding, for whatever it is worth
On 31 August 2026 the UK government opened a £100m competition backing British AI companies working on public service problems. Eligibility detail was still emerging at announcement. It is not a route for most SMEs directly, but it signals continued public appetite for adoption, and it is relevant to any firm whose customers include the public sector.
What we would actually do this quarter
- Pick one workflow that is high volume, rule shaped at the edges and cheap to get wrong.
- Write down the failure modes before the build brief. If you cannot list them, the scope is too wide.
- Assemble an evaluation set from 50 to 100 of your own historical cases and score against it.
- Set a monthly spend ceiling and instrument usage from the first live call or message.
- Name one owner with authority to change the process, not a steering group.
- Re-measure at 90 days against the baseline you took before anything changed.
The firms getting return are not the ones that moved first. They are the ones that scoped smallest.
The gap McKinsey measured is real, and it will keep widening while smaller firms treat agents as a procurement decision rather than an operating one. The good news in the same week's data is that closing it does not require enterprise budget. It requires one workflow, one owner and a test set.
Frequently asked questions
- Are AI agents worth it for a business under 250 people?
- For a single high volume workflow, yes, and the August 2026 Salesforce data suggests payback around 6.5 months where scope is narrow and escalation paths are clear. Across five workflows at once, usually not, because none of them gets the evaluation and ownership that makes an agent reliable.
- How much does a first AI agent cost to run now?
- Usage based pricing from the major providers, including Google's pay as you go option announced on 26 August 2026, means the running cost tracks volume rather than seats. Budget for build, model and infrastructure usage, and a monthly operating fee for monitoring and fixes. Agree a ceiling before launch.
- Does the EU AI Act apply to a UK SME?
- It can, if you place AI systems on the EU market or their output is used in the EU. Transparency obligations are enforceable now, with fines up to 3 percent of global turnover, while high risk obligations were deferred to December 2027 by the AI Omnibus regulation that took force on 27 July 2026.
- Should we wait for the technology to settle?
- No, but the alternative to waiting is not a broad rollout. Run one narrowly scoped workflow with a real test set. That gives you a measured result and an internal capability regardless of which model leads next year.
Work through this with us
brep runs AI operating reviews, workforce training and agent deployment for businesses of 10 to 250 people in the UK and US. One diagnosis, one workflow, someone accountable for keeping it working.