Most enquiry problems are solved by better form fields and a few plain rules. Add an AI step when someone has to read free text before anything else can happen. Consider an AI agent only when the next step genuinely depends on what it finds, and keep anything that commits your business behind a person.
That order matters because each step up costs more to run, test and maintain. A rule does the same thing every time and can be read in an audit. A model can give a different answer to the same message, and it can be confidently wrong. Anthropic, which builds the Claude models, puts it plainly in its guidance on building agents: find the simplest solution possible and only increase complexity when needed, which might mean not building an agent at all.
It also helps to be precise about words. In a workflow, the steps are fixed in advance and a model may do one job inside them. In an agent, the model decides which steps to take and which tools to use. Plenty of useful “AI enquiry handling” is really a workflow with one AI step, and that is usually a good thing.
Four options, from simplest to most complex
| Option | What it does | Good fit | Watch for |
|---|---|---|---|
| 1. Fix the form | Required fields, a service dropdown, postcode, budget band, target date | Enquiries arrive missing the same two or three details | Too many required fields put people off; ask only what changes your reply |
| 2. Plain rules | Route by service field, flag postcodes outside your area, catch duplicates, send an acknowledgement | The decision can be written as “if this, then that” | Rules multiply; review them when services or areas change |
| 3. One AI step in a fixed workflow | Reads a free-text message into fields, lists missing details, drafts a reply for review | The bottleneck is reading messages, not deciding what to do next | Output must be checked; the workflow, not the model, decides what happens next |
| 4. A bounded agent | Chooses which read-only lookups to make (past enquiries, service notes, calendar) and prepares a review packet | The route changes depending on what it finds across several sources | Higher running cost, a test set to maintain, and every action still needs limits |
Option 3 is where many businesses get most of the benefit. Our guide to using an AI agent to qualify website enquiries walks through that kind of draft-only intake packet in detail.
Ask five questions about your enquiries
- Is the information already in a structured field? If you route by the service someone picked and their postcode, that is a rule. You do not need a model to read a dropdown.
- Would two of your team route it the same way from a written rule? If yes, write the rule. If they would argue about it, the decision involves judgement and may need a person or a model to read the message.
- Is reading the message the slow part? Long, vague or mixed messages (“new website, maybe ads too, need it before our open day”) are where one AI step earns its place: extracting needs, dates and missing details.
- Does the next lookup depend on what the first one finds? An existing customer asking about an invoice and new work in one email may need the account checked before anyone knows the route. That is the territory where an agent can help.
- What happens if it gets one wrong? If a wrong answer means a missed lead or an awkward promise, keep the decision and any customer message with a person, whichever option you choose.
Sort a sample of real enquiries before you build
Take your last 30 to 50 enquiries, with names and contact details removed. Label each one with the simplest option that would have handled it: form or rule, needed reading, needed investigation across records, or needed a person’s judgement from the start. This takes an hour or two and is cheaper than building the wrong thing.
An illustrative sample of 40 might sort like this: 26 could be routed from the service and postcode fields, 9 were free-text messages that needed reading to work out what the person wanted, 3 needed someone to check whether they were an existing customer, and 2 were complaints that should go straight to a person. These numbers are made up to show the method, not a benchmark.
In that example, the answer is not an agent. It is a form tweak and routing rules for most enquiries, one AI step to turn the nine free-text messages into a short summary for review, and a clear rule that complaints skip automation altogether. An agent would only be worth testing if the “check the account” cases grew or became time-consuming.
Three worked examples
- A heating engineer covers twelve postcode districts. Enquiries from elsewhere waste time. Answer: a postcode rule that flags out-of-area enquiries and a polite, pre-approved reply that a person sends. No AI needed.
- A small agency gets messages such as “We need a new site and help with Google Ads, budget not sure, launching in spring.” Answer: one AI step that extracts the services mentioned, the timing and the missing details (budget range, current site, who decides), then rules route the summary to the right person, who replies.
- A business with existing customers gets emails mixing support questions, new work and billing. The right route depends on the customer’s account, open jobs and past invoices. Answer: possibly a bounded agent that can read, not change, those records and prepares a packet saying what it found and what it suggests. A person still decides and replies.
All three are illustrative. The point is that the decision comes from the shape of the work, not from wanting to use AI.
Keep rules around any agent you do build
Even when an agent is the right choice, the things that must always be true should stay in plain code or system settings: who it can contact, which records it can read, which actions need approval, how duplicates are caught and how long data is kept. A prompt that says “never email a customer without approval” is an instruction, not a control.
The OWASP guidance on excessive agency in AI applications makes the same point: authorisation belongs in the downstream systems, not in the model’s judgement, and high-impact actions should need a person’s approval. Give the agent the fewest tools and the narrowest permissions that do the job.
Count the running costs honestly too. An agent needs model calls on every enquiry, a set of test cases you rerun when anything changes, and someone who reviews its output. A rule needs editing when a service or area changes. If the agent does not save more time than those costs take, the plainer option wins.
Where to go next
If your enquiries mostly arrive complete, look at the form and the routing first: our guides to connecting your website to a CRM and choosing between Zapier and a custom integration cover that side. If reading messages is the bottleneck, start with one AI step and a person approving replies. Bring a sample of your enquiries to Foundry and we can help you decide whether you need a form change, an integration or a narrow internal agent.
