Lana Korzhuk — Founder & CEO of SIMARA AI

Lana Korzhuk

Founder & CEO · LinkedIn

With 20+ years progressing from senior developer to Chief Operating Officer, Lana brings deep expertise in IT systems, ERP implementation, and operational strategy.

Published · 12 min read

AI Agents for Small Business: When to Buy, Build or Use Rules Instead

AI Agents for Small Business: When to Buy, Build or Use Rules Instead
💡

TL;DR

  • ●Choose rules when job requests contain the fields needed for routing, and a person can handle exceptions without creating an unacceptable backlog.
  • ●Shortlist a packaged AI agent when requests vary in wording and a specific product can demonstrate the access, review controls and ownership arrangement you need.
  • ●Consider a custom agent only when a material requirement cannot be met by rules or a suitably configured package, and you can support the integration and ongoing maintenance.

For an incoming job request, decide whether the ability to interpret the request and select an action changes the outcome enough to justify buying and owning an agent.

We recommend comparing all three approaches against the same request, systems and human review boundary. Rules may be sufficient, even if an agent gives a more polished demonstration. A packaged agent, meanwhile, is not necessarily suitable just because it can summarise an email.

This article gives you a selection worksheet for that decision. For the wider landscape, including tools and financial justification, see our workflow automation guide for UK SMEs.

The contenders overview

The workflow in this comparison receives a job request, identifies its service type and location, and places it with the appropriate operations team. A person decides how to handle unclear or exceptional requests. The options differ in how they interpret the request and how much of the solution you need to own.

Definition: Rules-based automation: a workflow that follows predefined conditions, such as routing a request with a selected service type to a named team.

Rules instead of an agent. Use structured fields and explicit routing conditions. Requests that do not match a rule go to a manual queue. Start here when you can consistently capture the information needed for routing.

Buy a packaged agent. Assess an existing product configured for the workflow. It may offer language-model interpretation and actions, but the supplier must demonstrate the exact access, controls and permitted configuration for the product and plan you are considering.

Build a custom agent. Commission a solution tailored to your systems and decision boundaries. Custom development can address a specific gap, but you will still need to agree who maintains the solution and its dependencies.

Definition: AI agent: a system that interprets a request and can select or invoke tools to carry out permitted tasks. The tools and authority it receives determine what it can do.

Microsoft's architecture guidance distinguishes an agent that orchestrates requests from the individual tools it invokes. Ask which actions need to be selected dynamically, rather than assuming every workflow involving AI needs an agent.

Use-case comparison: the same job-request worksheet

Before seeing a demonstration, complete this worksheet using your actual job-request process. The entries are decision criteria, not claims about any particular product.

Selection factor Rules-based workflow Packaged agent Custom agent
Request variability A strong fit when required fields and routes are explicit; send unmatched requests to a person Consider when varied wording needs interpretation and the package can show acceptable handling of your examples Consider when the interpretation or action sequence is specific enough that a package cannot meet the requirement
System access Specify the fields and destinations the rules need Verify the exact records, permissions and actions available in your configuration Specify each interface and permission to be built, tested and maintained
Human review A person resolves missing fields and exceptions Check that uncertain interpretations can be reviewed before the chosen action Design and test the review boundary; custom code does not make review automatic
Ownership costs Rule changes, exception handling and platform administration Subscription or usage charges, configuration, review and supplier dependency Development, hosting or usage, support, changes and review, plus any external services

Work through the rows in order. Define the minimum acceptable routing outcome first. Then eliminate options that cannot reach the required information or respect the review boundary. Compare the remaining ownership commitments only after that. A lower quote does not make up for a missing control.

Definition: Human review boundary: the point at which a named person must inspect or decide something before the workflow can proceed. It should specify the decision, not merely say that someone can monitor the system.

If you use Microsoft 365 and are considering Power Automate for the rules route, treat it as a candidate to investigate. Do not assume your licence includes the required connectors or permissions. Ask for a demonstration in your own configuration.

Pricing: compare the ownership commitment, not the headline fee

A useful comparison includes the work your team retains as well as the supplier's quote. Request the same scope from each shortlisted option: receiving the job request, identifying the route, presenting exceptions for review and maintaining the agreed configuration.

For rules, record configuration and change costs, platform charges where applicable, and the staff time needed to handle unmatched requests. For a packaged agent, ask for implementation and recurring charges, any usage-dependent costs, support terms, and who changes instructions or permissions. For a custom agent, request the build quote along with expected costs for hosting or external services, support, testing changes and transferring ownership.

Definition: Ownership cost: the combined supplier and internal effort needed to operate, review, maintain and eventually change or replace a workflow.

Do not treat a supplier's estimate of reduced review effort as an established saving. Compare it with a trial using your request mix. If you need the separate financial case for the chosen option, use the AI ROI analysis guide rather than treating the selection worksheet as a payback calculation.

If you first need to understand where your business is losing profit, the optional Value Driver Tree can help you identify workflows to investigate. It does not choose an agent for this request process.

Scaling: what changes as the request mix grows?

Growing volume alone does not establish that an agent is necessary. Look instead at whether the mix of requests and exceptions has changed enough to make the current route difficult to maintain.

With rules, watch for an expanding set of conditions that staff struggle to understand or update. If the request form remains stable and exceptions remain manageable, higher volume may still favour the simpler route. If customers increasingly send free-text requests, consider whether changing the intake form would solve the problem before buying interpretation capability.

With a package, ask how configuration changes are made, who can review them and what happens if a necessary integration or permission changes. With a custom agent, agree who owns the code, system accounts, operational documentation and support when the original developer is unavailable. Resolve these questions before commissioning rather than assuming every supplier provides the same arrangements.

Microsoft's agent orchestration guidance notes that multiagent coordination adds overhead and failure modes. For this job-request decision, start with the least complex design that meets the requirement. Multiple agents do not remove the need to demonstrate that one route works.

Worked comparison: which route would you shortlist?

The following scenarios are illustrative and synthetic. The request IDs, system arrangements and review rules are invented assumptions, not observed outcomes or vendor capabilities. Each scenario changes the same incoming job-request workflow. Record actual trial results separately before purchasing.

Scenario A: structured requests, rules are sufficient

Illustrative input: Request JOB-A arrives through a form with service type repair, location South and a contact field. The business has an agreed rule: repair requests for South go to the South operations queue. Missing service type or location goes to manual review.

  1. Rules: Check the selected fields and route to South operations. If a required field is absent, hold the request for a person.
  2. Packaged agent: Ask what additional work it would remove. Interpreting free text adds little to this request because the required fields are already present.
  3. Custom agent: There is no stated requirement here that justifies custom interpretation or dynamic tool selection.

Shortlist: Rules, provided the form is usable and the exception queue is acceptable to the owner. There is not yet a reason to take on the extra ownership burden of a more elaborate solution.

Scenario B: varied wording, test a package

Illustrative input: Request JOB-B arrives by email: “The front entrance at our South site needs a repair.” Assume the business cannot require the sender to use a structured form. The desired output is a proposed repair category and South location for an operations employee to check before routing.

  1. Rules: A fixed phrase list might catch the example, but the buyer must decide how to handle alternative wording and unrecognised requests. Manual review remains the fallback.
  2. Packaged agent: Ask a supplier to show the proposed category, location and source text on this same request, with the employee retaining the routing decision. Verify that this is possible in the offered configuration.
  3. Custom agent: Do not commission one merely because the input is an email. First establish whether the package meets the required review and access conditions.

Shortlist: A package alongside rules with manual review. Select the package only if its demonstrated reduction in handling effort and its ownership terms justify the difference. A fluent summary is not enough without a usable review step.

Scenario C: a specific access gap may justify custom work

Illustrative input: Request JOB-C says a repair is needed at the South site. Assume the operations employee must also see a site-access restriction held in an internal system before assigning the job. The restriction is visible only through an interface the business is authorised to use.

  1. Rules: Compare a route that sends the request to a person, who checks the restriction manually. This remains the baseline.
  2. Packaged agent: Ask the supplier to demonstrate access to the specific restriction, under the intended account, and present it for review. A generic integration claim is not evidence that it can meet this requirement.
  3. Custom agent: Request a scoped design and quote only if the manual check is materially burdensome, the package cannot meet the requirement, and the authorised interface can support the proposed access. The person still decides the assignment.

Conditional shortlist: Custom development may be justified if a credible design addresses the verified access gap and its ownership commitment is acceptable. If a configured package or a simple integration can supply the restriction safely, compare that simpler option first. If the interface is unavailable or unauthorised, keep the check manual rather than promising that a custom agent will solve it.

What evidence should change the shortlist?

Turn the scenarios into a buyer's record before asking for proposals. For each representative request, write down the information available, the desired output, the permitted action and the person who handles uncertainty. Include routine requests, ambiguous wording and requests with information missing. Use synthetic copies where appropriate.

For each contender, record expected behaviour, observed behaviour, the evidence shown and the remaining manual work. Do not mark an option as proven because a supplier says it is configurable. If the package proposes a category but cannot show the reviewer the source or the relevant restriction, for example, record that specific gap and ask whether it can be resolved in the quoted scope.

Custom development moves up the shortlist when all of the following are supported: the unmet requirement matters commercially, simpler alternatives have been investigated, the necessary access is authorised and feasible, and an owner can accept the build and support commitment. If any point is unresolved, seek evidence or retain the manual step.

This is a selection exercise, not proof that a production integration is safe. Once an option involves moving an approved record between systems, use the separate cross-system handoff acceptance guide for its detailed tests of approvals, duplicate events, failures and recovery.

Trade-offs, risks and limitations

Rules make the routing logic explicit, but someone must maintain it as service categories or territories change. They can also move work into a manual queue rather than remove it. Inspect that queue before concluding that rules are cheaper to operate.

A packaged agent may reduce configuration work, but the buyer can become dependent on the supplier's supported actions, commercial terms and change process. A product that interprets requests well may still leave staff copying information between systems or checking every output. Ask who performs those tasks and include them in the comparison.

A custom agent gives you the opportunity to specify a narrower fit, but it does not guarantee correct interpretation or reliable access. Its burden includes support, changes to connected systems, ongoing review and a route to take over from the developer. More control also means more responsibility for specifying and maintaining the solution.

For any option that handles personal data, ask what information it accesses, where that information is processed, who can see it and how it is retained or removed. Obtain advice appropriate to your circumstances rather than treating an agent label or supplier assurance as proof of GDPR compliance.

When this recommendation does not apply

Do not use this worksheet to award a contract when the evidence only shows a generic demonstration. A supplier's sample may omit your unusual wording, unavailable data or required review step. Re-run the comparison with representative or synthetic requests and your stated boundaries.

The three-way choice also breaks down if the job cannot be safely routed from the information available. For example, if the site restriction in JOB-C is unavailable to both staff and the proposed solution, neither a packaged nor a custom agent can make the missing evidence reliable. Keep the decision with an authorised person and resolve the information gap first.

If the package passes the interpretation demonstration but cannot access the required record, it is not yet a fit for the complete workflow. Compare a manual check or a separate integration. If a custom supplier can build the access but cannot identify who will support it, keep custom development off the shortlist until that ownership issue is resolved.

Final verdict: choose the simplest defensible route

For a structured incoming job request, start with rules and an owned exception queue. For varied requests, compare a packaged agent against rules plus manual review using the same inputs and the same permitted outcome. Commission a custom agent only where a verified, material gap remains and the business is willing to own the resulting solution.

Choose based on the work the system must do, rather than assuming custom development is inherently more advanced or more valuable. Keep interpretation, system access and authority separate: an option can be good at understanding a request without being suitable to act on it.

What to explore next

Sources & Further Reading

Frequently Asked Questions

Common questions about this topic, answered briefly.

Do AI agents for small business always need to take actions?
An agent can interpret a request and invoke permitted tools, but the buyer should define which actions, if any, it may take without a person's approval. A proposed route for human review may be sufficient for the job-request workflow.
When should I build AI agents for business rather than buy one?
Consider custom development when a material requirement remains unmet after comparing rules, manual review and a suitably configured package. Verify the necessary system access and agree who will maintain the solution before commissioning it.
Can rules and AI agents work together?
Yes. For example, an agent could propose a category from free text while a predefined rule limits what happens next. The chosen solution still needs to demonstrate the exact design and controls.
How do I compare AI agents for your business without a product list?
Give each shortlisted supplier the same representative request, desired output, access requirement and review boundary. Record what each demonstrates, what remains manual and the full ownership commitment. Eliminate options that cannot meet mandatory requirements before comparing their quotes. [Contact SIMARA AI](/contact) to discuss the decision for one of your workflows.

See your numbers before you spend a penny

Free, no sign-up: the Value Driver Tree maps your value drivers in about two minutes and shows which workflows drain profit and where automation has the highest leverage.

Try the free Value Driver Tree

Or talk to us directly: