AI agents for business: what they do, what they cost

What AI agents for business actually are

AI agents for business are systems that take multi-step action on their own. They differ from single-turn AI automation in a specific way. An agent can take a sequence of actions, observe the result of each one, and adjust its next step based on what it found. A single-turn automation reads an email and drafts a reply, and that is the whole job. An AI agent reads the email, checks the availability calendar, looks up the customer's booking history, composes the reply, and queues it for approval. The agent completes the full workflow, including the intermediate lookups, without a human doing each step by hand. In 2026 these agents are moving from experimental to production-ready, but only for a narrow set of high-frequency, well-defined workflows at small and mid-sized businesses. On those workflows the return is real.

The difference between AI automation and an AI agent

Single-turn AI automation works in one move. Input comes in, the AI processes it, output goes out. An AI agent works in several moves. Input comes in, the agent takes action one, observes the result, takes action two based on what it found, and produces a final output once it has enough to act on.

That difference drives both cost and reliability. Single-turn automation is cheaper to build and more predictable, because there is only one path through it. Agents are more capable but harder to build and monitor, because each intermediate step is another place something can go wrong. For most businesses the right starting point is single-turn automation on the highest-frequency workflows, and agents only where the intermediate lookups are themselves worth automating. If you want the foundation first, here is what AI automation is and how it sits underneath agent work.

What AI agent workflows work reliably for SMEs in 2026

The workflows where agents earn their cost over single-turn automation are the ones that need a business to check several data sources before producing an output. Here are three that hold up in practice.

Multi-step lead qualification. A lead arrives. The agent checks whether this email or phone number has contacted the business before, looks up the customer record in the CRM, and checks for open tickets or notes. Then it asks the qualifying questions, adapted to what it already knows. Single-turn automation cannot do the lookup step. An agent can, and it stops asking a returning customer questions you already have answers to.

Booking management with conflict checking. An agent reads a reservation inquiry, checks the booking calendar for availability, checks room or table inventory, and checks for existing bookings from the same customer before drafting the confirmation. A London hospitality group had this running across eight venues, handling multi-venue availability checking that previously required a manager to check three separate systems by hand. The agent collapses that into one pass.

Candidate pipeline management. An agent checks a candidate record, looks at their LinkedIn activity in the last seven days, checks whether anyone at the firm has contacted them in the last 30 days, and then drafts outreach matched to their current situation. Single-turn automation would send the same message to every candidate regardless of recent activity.

What AI agents can and cannot do today

The real capabilities in 2026 are narrower than vendor marketing suggests and wider than most owners realise. Agents reliably read and classify unstructured inputs. An agent can read inbound emails, WhatsApp messages, chat transcripts, and reviews and sort them by intent, urgency, and the type of response required, with accuracy above 90 percent in most business contexts. They draft contextual responses that need only a 30-second human review before sending, not a template but an actual reply built from your pricing, policies, and prior interactions. They handle sequential multi-step tasks: book an appointment, send a confirmation, add a CRM task, and update the deal stage, all from one customer action. And they turn a brief plus a set of documents into a structured summary in minutes.

What they cannot do reliably is the other half of the picture. Keep a human in the loop on high-stakes decisions. An agent should not make binding financial, legal, or clinical calls on its own, because the error rate, while low, is not low enough when one mistake is expensive. Agents also cannot touch the physical world: they operate on data and text. And they are poor at building trust from scratch. They strengthen existing relationships through consistent communication, but struggle where a buyer is evaluating a person, as in high-value enterprise sales or professional services.

How AI agents differ from automation tools

Automation tools like Zapier, Make, and n8n move data between systems on predefined rules. If this, then that. They are excellent for high-volume, low-variance tasks where the logic is clear and exceptions are rare. AI agents operate on goals rather than rules. You give an agent an objective, for example "follow up with this lead until they book a call or explicitly decline," and it decides how to pursue it from context: reading previous interactions, picking the right channel, adapting based on what has worked, and knowing when to stop. Rules-based automation breaks the moment reality deviates from the script. Handling deviation is exactly what an agent is for, which makes agents most valuable where the decision space is too wide to script every path.

What do AI agents cost for small businesses

The infrastructure cost runs higher than single-turn automation because agents make more API calls per run. A single-turn automation that costs £0.02 per run might cost £0.08 to £0.15 as an agent, which at SME volumes is roughly £20 to £80 a month in AI API costs. The build cost is also higher. A single-turn automation takes one to two weeks to build and test. An agent takes two to three weeks, because each intermediate step has to be tested on its own and in combination, and the edge cases are messier.

On deployment, price tracks volume and decision complexity. A basic agent handling email classification and response drafting for a team of 10 runs about £200 to £500 a month including API costs, typically built on OpenAI's API or Anthropic's Claude API with a thin layer of business logic. A more sophisticated agent handling multi-channel engagement across email, WhatsApp, and live chat with CRM integration runs £800 to £2,500 a month. Custom agents for a specific high-value workflow, such as lead qualification for a professional services firm, patient triage for a clinic, or candidate screening for a recruiter, are usually priced as projects at £5,000 to £20,000 to build, with ongoing API costs of £200 to £1,000 a month.

The question that decides any agent investment is simple. What is this workflow costing you now, and what would faster or higher-quality execution be worth? If the answer is not at least three times the monthly cost of the agent, the case is weak. For the full picture, see how much AI automation costs.

Should you start with agents or single-turn automation

Start with single-turn. Build a lead qualifier, a booking confirmation drafter, or a CRM reconciler as a single-turn automation first. Run it for 60 days. Learn which edge cases come up. Then decide whether an agent would handle those edge cases better than adjusting the single-turn logic would.

The most common mistake is designing an agent when a single-turn automation would have done the job at a third of the cost and half the build time. The second is building an agent before the data underneath the workflow is clean. An agent navigating dirty data produces confident wrong answers at speed, which is worse than no automation at all. Fix the data first. For more on sequencing, see AI for business process automation and the broader case for AI automation for business.

How twohundred would approach an AI agent build

In practice, the order matters more than the tooling. twohundred starts by picking one high-frequency workflow, measuring what it currently costs in time and missed revenue, and building the smallest version that produces a measurable output: a response time, a resolution rate, a conversion number. We build single-turn first wherever it will do, because it is cheaper to run and easier to trust, and we only upgrade to a full agent when the intermediate lookups genuinely need a machine deciding the next step. We clean the underlying data before anything reads it, and we keep a human in the approval loop on anything with real downside. If you want a workflow scoped this way, with the costs and the return laid out before a line of code, that is the work we do under AI workflow automation.

Frequently asked questions

What is the difference between an AI agent and a chatbot?

A chatbot responds to one message at a time with no memory of what it did two steps ago. An AI agent pursues a goal across several steps, checking data sources and adjusting as it goes. A chatbot can answer "what are your opening hours." An agent can read a booking request, check availability across systems, and draft a confirmation. The agent does the lookups in between, which a chatbot cannot.

How much does an AI agent cost for a small business?

A basic agent handling email classification and drafting for a team of 10 runs about £200 to £500 a month including API costs. Multi-channel agents with CRM integration run £800 to £2,500 a month. Custom agents for a specific high-value workflow are usually priced as a project at £5,000 to £20,000 to build, plus £200 to £1,000 a month in API costs. The right spend depends on what the workflow is worth to you.

Are AI agents reliable enough to run without supervision?

For routine, well-defined tasks, yes, agents classify and draft with accuracy above 90 percent in most business contexts. For high-stakes decisions, no. Keep a human in the approval loop on anything financial, legal, or clinical, where one error is expensive. The sensible pattern is autonomy on low-risk volume and human review on the decisions that carry real downside.

Should an SME start with an AI agent or simpler automation?

Start with single-turn automation on your highest-frequency workflow and run it for 60 days. Most businesses find a single-turn build handles the job at a third of the cost of an agent. Only upgrade to an agent when the edge cases genuinely need a machine deciding each next step. Building an agent before the underlying data is clean is the most expensive mistake here.

If you want help scoping an AI agent for a specific workflow in your business, book a session with our team.

---

Related Services

For businesses looking to build autonomous workflows, AI agent development covers the architecture and delivery end to end. Once agents are in place, AI implementation services handles the deployment and rollout phase.