AI agents in 2026: what's real in B2B operations

In 2026, AI agents do bounded, checkable work well: research, enrichment, drafting, classification and triage, with a human or a programmatic check reviewing output before it matters. What they do not yet do reliably is run multi-step business processes unsupervised — most "fully autonomous" demonstrations still depend on tidy inputs that real operations rarely supply. The practical question is not whether agents work, but which of your tasks tolerate their failure modes.

What is an agent, once the marketing is removed?

An agent is a language model given tools and a goal, allowed to take multiple steps — search, read, write, call an API — and to decide its own next action. That looping autonomy is the whole difference from a chat prompt, and it cuts both ways: each step's output feeds the next step, so small errors can compound quietly across a run.

This is why I hold the same line here as in the broader reality check on AI automation for B2B firms: judge the mechanism, not the demo. A demo shows one successful path through the loop. Operations expose every other path.

What do agents genuinely do well in B2B operations today?

The reliable uses share a shape: the task is bounded, the output is verifiable, and something checks it before consequences follow. In our own work and our clients' operations, that commonly means:

• Lead and account research. Reading company websites, filings and public profiles, then returning structured summaries against a fixed brief. We run this daily — the build log of an agent that builds and enriches a B2B database shows what that looks like in practice, verification steps included.
• Enrichment and normalisation. Turning scraped or exported mess into consistent fields: job titles mapped to seniority, addresses standardised, duplicates flagged for review.
• Drafting. First passes at internal documents, meeting summaries, and reply drafts a human edits before sending.
• Triage. Classifying inbound — enquiry, support, invoice query, spam — and routing it, with low-confidence cases escalated to a person.

None of this is glamorous. All of it removes hours, and the error rate is survivable because a check sits between the agent and the outcome.

Where does it remain demo-ware?

In my experience, the gap between demonstration and deployment is widest in four places. Unsupervised end-to-end processes with real consequences — sending client communications, moving money, committing to dates — still fail in ways that surface at the worst time. Long-horizon tasks that run for hours or days tend to drift off-brief without checkpoints. Agents asked to exercise your firm's judgement fail predictably when that judgement was never written down for them. And agents driving a browser across arbitrary third-party sites remain brittle: when the page layout changes, then the run breaks, usually silently.

None of these are permanent verdicts — the ground has moved every year I have worked with this. They are the current boundary, stated plainly so you can plan around it rather than discover it.

How do you deploy an agent without betting the process on it?

The mechanism: when a task passes the bounded-and-verifiable test, then give the agent the narrowest tool set that can complete it — read access before write access, one system before five. When the agent produces output, then route it into a review queue rather than into the live system. When a fortnight of review shows corrections settling into a pattern, then automate those specific checks and widen the agent's scope one notch — not three. And before any of this touches client information, then settle your boundary rules for client data and AI tools, because retrofitting a privacy position after connecting systems is far harder than setting one first.

Run that loop and an agent earns autonomy the way a new hire does: gradually, against evidence.

What should a five-to-fifty-person firm actually do this year?

Treat agents as fast, tireless junior staff with no memory of your business beyond what you hand them, and no shame about being wrong. That framing gets the economics right — real leverage, real supervision cost — without either dismissing the technology or trusting it with the keys.

Start with one contained project, chosen with the same discipline as picking your first automation project: frequent, rule-based, small blast radius. Prove it, measure it, then extend. The firms getting value from agents in 2026 are not the ones running the most ambitious pilots; they are the ones that shipped something narrow, kept the checks, and compounded from there.

---

Next step: the Growth System Audit — £450, seven days, credited against any build — maps where your growth system leaks and what to build first.