What you hold at the end is a folder a bank manager could follow without you in the room: one workflow named in a single sentence, the systems it may touch listed with the access granted to each, last week's tool calls exported and readable line by line, a cost per completed action, and one named person who answers the phone when a run stops at two in the morning.
The example is a twelve-person wholesaler with a two-person credit desk. The version of agentic AI that survives contact with that business is not a chat window bolted onto the website. It is an agent that reads an inbound order email, checks stock in the ERP, applies the customer's agreed price from the CRM, drafts the confirmation, and stops dead when the order value crosses the credit limit or the account carries an invoice more than 30 days past due. Four systems, one job, two stop conditions. Everything below is about getting that to its second month.
The shift matters now because the economics have changed shape. Salesforce's State of Agentic AI in the Enterprise report, which surveyed 2,025 agentic AI decision-makers across 20 countries including 21 per cent from small and medium businesses, puts average time to return on investment at about eight months for organisations with agents in production. Eight months sits inside the planning horizon of a firm with no chief technology officer.
An agent picks its own route, and that route is the risk
An agent is a planner with a toolbox, and the toolbox is the part you have to secure. GOV.UK's AI Insights guidance on agentic AI makes the distinction operational: conventional software follows a fixed pathway, while an agentic system receives a set of capabilities, each described in natural language, and chooses which to combine in order to reach an objective. A chatbot returns text. An agent picks up a tool and uses it.
That is also why testing an agent is not testing software in the ordinary sense. The same guidance flags the failure that catches teams most often, which is moving a working proof of concept straight into production without the profiling and non-functional work a live system needs. It warns too that the frameworks change quickly and break compatibility between versions, so every version has to be pinned and recorded. It gives a small example worth memorising: a pricing tool described inaccurately in its natural-language definition will be selected for the wrong job, and the system will do exactly what it was told.
The practical consequence is that the audit question is not what the model said. It is which tools the agent called, in what order, with which arguments, and what came back. Log that from the first day, because it cannot be reconstructed later.
Bound one job before you buy anything
Pick one workflow with a clear start and a clear stop, and clean the data that workflow touches. Salesforce's survey found the two strongest signals of a fast rollout tied at 36 per cent each: clean, accessible data at the point where the agent acts, and a tightly bounded use case. In the same survey, only 31 per cent had fully unified their data before deployment and 69 per cent of the remainder still reached return on investment, which is not permission to ignore data quality: 63 per cent of organisations hit agent errors caused by stale, missing or wrong data.
The NCSC's guidance on adopting agentic AI says the same thing in security language. Start with tightly bounded pilots on clearly defined, low-risk tasks, build confidence, then widen the scope. It opens with "make sure you can walk before you run", and the phrase is doing real work. It also asks a question most vendors will not ask on your behalf: whether AI is needed at all, or whether the process could be simplified, removed or automated in a lower-risk way first.
For the wholesaler, that means the agent's first job is drafting order confirmations for accounts comfortably inside their credit limit. It does not touch the credit decision. It does not touch pricing exceptions. It does not send anything to a customer whose account is on hold.
Autonomy is not the absence of a reviewer
Increased system autonomy has so far necessitated more human re-entry. In the Salesforce report, 81 per cent of organisations reported more human involvement in workflows as their AI deployments expanded, and 89 per cent said they were rethinking workforce models to account for agents. Read that as evidence that the work moves: routine questions and out-of-hours coverage go to the agent, and the relationship moments come back to the team.
The NCSC is blunt about where accountability sits. A system may take an action, but humans remain accountable for the decision to deploy it, the access it was granted, the safeguards around it and the consequences of its operation. Four names belong on the page before the agent touches live data:
- who owns the agent
- who approves its access
- who monitors its behaviour
- who can stop it
If any line is blank, the deployment is not ready. GOV.UK adds the operational test that settles most arguments: if you cannot understand, monitor or contain what the agent does, it is not ready for production. For the wholesaler that translates into one escalation rule written in plain English, an order above the credit limit goes to the credit desk with the draft attached and nothing sent to the customer until a human releases it.
Governance bought at the start is cheaper than governance retrofitted
The organisations with stronger governance took longer to reach return on investment, at 9.3 months against 7.2 months for those with weaker governance, according to the Salesforce report. They also caught problems proactively 68 per cent of the time against 55 per cent. That trade is worth taking, and the regret data explains why: of the 43 per cent of organisations that had to stall or scale back an initiative, 39 per cent said their top regret was not budgeting for AI fluency training and reskilling up front, and 38 per cent wished they had set clearer governance before scaling, according to the State of Agentic AI Report for Small Business.
For a small firm, governance is not a committee and not a policy document with a review cycle. It is one page that says who reviews what the agent produces, what threshold forces an escalation, how long the logs are kept, and what the agent is not permitted to do under any circumstances. Write it before the pilot. Retrofitting it after an incident costs more than the delay of writing it first, which is the whole point of that 9.3-month figure.
Price the action, not the seat
An agent's cost moves with how much work it does, which is why seat-based budgeting fails. Zuora's guide to pricing agentic AI sets out four models in use. Per agent behaves like a seat licence and holds only while scope stays narrow. Per activity charges for countable steps, the shape Salesforce uses for Agentforce at $2 per conversation, as detailed in Pricing Agentic AI: A Practical Guide. Per output charges per artefact produced. Per outcome charges for a defined result, which is how Intercom prices its Fin agent at $0.99 per resolved ticket, as detailed in Pricing Agentic AI: A Practical Guide. Outcome pricing transfers cost variance to the seller and demands that both sides trust the same measurement, so it is the model that rewards a clear definition of done.
The build is not free either. Sparkout's 2026 cost breakdown puts a proof of concept at $15,000 to $35,000, a minimum viable product at $25,000 to $60,000, and a production system from $60,000 upward, in dollars. Treat those as vendor figures and check them against your own quotes, but the order of magnitude is useful the next time someone proposes a six-figure platform before a single workflow has been bounded.
Budget the running cost per completed action. If one agent processes thousands of actions a day, a flat fee per seat will eventually either bankrupt the vendor or cap what you can do with the system.
The order to do it in
Write the one-sentence job description. List the systems it needs and the access level for each. Name the four people. Decide where it stops and who it hands to. Then buy, build or configure only that.
The eight-month return in the Salesforce data is an average across organisations that did the boring parts first. The firms that skipped them are inside the 43 per cent who had to stop and start again, and they paid twice for the same lesson.
Sources
- Takeaways from the State of Agentic AI Report for SMB
- Thinking carefully before adopting agentic AI
- AI Insights: Agentic AI
- Pricing Agentic AI: A Practical Guide
- AI Agent Development Cost in 2026
Source: 4 Takeaways From the State of Agentic AI Report for Small Business, Salesforce





