A 40-person e-commerce business sat down to plan its first serious AI deployment. The marketing director wanted a chatbot on the front of the website to deal with the rising support queue. The operations director pointed out that the firm processes around 600 supplier invoices a month, by hand, costing one full FTE about three days a week. Two completely different decisions, the same line in the budget.
Six months later, the chatbot was live, escalating four in every ten conversations to a human, and the support queue was bigger. The invoices were still being keyed in by hand. The opening question they should have asked was “which use case can fail without anyone outside the firm noticing”, and they had asked the wrong one.
The choice you’re facing
Workflow automation deploys AI to structured back-office processes, things like invoice processing, document classification, email triage and lead qualification. It runs through tools like Zapier, Make.com, n8n and Microsoft Power Automate. Chatbots deploy AI to user-facing conversational interfaces, things like customer support, internal helpdesks and knowledge-base search. They run on natural language, sit at the boundary of the business, and every output becomes part of the brand.
That distinction tells you where risk concentrates. A workflow automation that miscodes an invoice causes an internal exception that finance catches at month-end. A chatbot that hallucinates a refund policy ends up in front of a customer, on a screenshot, on social media, possibly at tribunal. Glean’s framing captures the split cleanly: workflow bots execute deterministic processes, chatbots interpret and respond to natural language. Same AI category in the budget, two different blast radii.
For most UK SMEs in 2026, the question is which one to ship first. The British Chambers of Commerce found in March 2026 that more than half of UK firms are now using AI, up from 35 per cent a year earlier, with the clearest measured impact sitting in back-office automation rather than customer-facing chat.
When workflow automation is the right answer
Workflow automation is the right answer when a process is repetitive, rule-based, touches multiple systems, and the time saved goes back to people who can do higher-value work. Invoice processing is the textbook case. Document triage for law firms and insurers is another. Email routing, lead qualification, payroll and HR queries through Slack or Teams bots all sit here. Quadient and similar AP tooling reports 60 to 80 per cent reductions in manual handling time.
The brand exposure is low. If a workflow misclassifies a supplier invoice, the firm catches it internally, fixes it, and adjusts the rules. There is no screenshot, no journalist, no tribunal. The failure mode is recoverable, and the cost of a mistake is bounded. That matters more than the headline efficiency number, because it is the thing that lets you actually ship.
Workflow automation also forces useful discipline. To ship invoice automation you have to deduplicate the supplier database. To ship document triage you have to tidy your taxonomy. That data-quality work is the foundation anything customer-facing has to stand on later. Firms that try a chatbot first end up doing this work in reverse, against a live user-facing surface, which is a much harder place to do it.
The technical bar is low enough now that a finance manager or operations lead can build the first workflow without a developer. Zapier, Make.com and n8n cover the orchestration. A working version typically takes weeks, not months, and the platform fee runs in the tens of pounds rather than the hundreds.
When a chatbot is the right answer
A chatbot is the right answer when customer-support volume is the binding constraint on the business, when the questions are repetitive and answerable from curated content, and when a clean human-escalation path exists for everything outside the deflection band. The classic case is a consumer brand with thousands of monthly tickets where a third are routine “where’s my order” queries. Internal helpdesks for IT and HR are a quieter case to start with.
Realistic deflection rates matter here. Vendor decks routinely quote 70 to 90 per cent. Independent benchmarks from Supportbench and similar 2026 sources put consumer-support deflection at 30 to 50 per cent and B2B deflection at 20 to 35 per cent, because B2B questions are more nuanced and lean on judgement. Build the business case against the lower band. If the numbers only work at 80 per cent deflection, the project is fragile.
Knowledge-base search is the other useful chatbot pattern, particularly internally. A retrieval-grounded bot sat over the employee handbook, IT documentation and internal wiki collapses minutes of intranet hunting into one question, and the failure mode is much milder than a wrong answer to a customer. Teams bots, Slack bots, Intercom configured for internal use and custom builds against OpenAI or Claude all sit in this band.
If you do go customer-facing, ship with a managed platform that gives you audit trails, escalation paths and PII handling by default. Intercom, Drift, Ada and Yellow.ai all do this. Custom GPTs and Copilot Studio are fine for internal pilots, but for anything regulated or brand-exposed the managed platform is worth the extra spend. Read our 12-question due-diligence list before you sign.
What it costs to get wrong
Chatbot failures are public, fast and quotable. Air Canada’s chatbot promised a bereavement-fare refund the airline tried to disown. The Civil Resolution Tribunal held the airline to its chatbot’s words, and the principle hardened: a chatbot’s output is a corporate statement, and the firm is liable. DPD’s chatbot was prompted into swearing at a customer in 2024. McDonald’s pulled its drive-through AI pilot in June 2024 after misorders went viral.
Lenovo’s customer-facing chatbot was breached by a 400-character prompt-injection attack in 2025 that exfiltrated session cookies, a clean illustration of what prompt injection actually is when it lands on a real product. Hallucination sits underneath the wider chatbot risk picture. When a customer-facing bot is not retrieval-grounded against verified content, it invents answers that sound plausible. The cost is paid in refunds, customer-service escalations, ICO attention under UK GDPR, and brand damage that does not come off easily.
Workflow automation failures are different and quieter, which makes them harder to spot. An invoice classifier mis-tags supplier invoices, the error propagates through the general ledger, and someone catches it at month-end reconciliation six weeks later. An approval workflow flags too many false positives, the approver starts rubber-stamping to clear the queue, and a control that looked like governance becomes theatre. A workflow that touches personal data without a human checkpoint walks straight into GDPR Article 22 territory, where solely-automated decisions about individuals carry rights and obligations the firm has to respect.
The blast-radius asymmetry is the heart of the case. A chatbot failure is loud and fast. A workflow-automation failure is quiet and slow. Both can be costly, but the firm has a much better shot at catching the quiet one before it leaves the building.
What to ask before you decide
Seven questions, in this order. One: where is the actual value, in customer-facing efficiency or in back-office capacity? Two: who is the end user, and what can safely fail in front of them? Three: what data does the system touch, and is it clean enough to act on? Four: what governance applies, including Article 22 for workflow automation that decides things about people, and ICO and FCA guidance for chatbots in regulated firms?
Five: build, buy or partner? For workflow automation, no-code platforms now make build viable for an operations lead. For customer-facing chatbots, buy a managed platform unless you have the engineering and governance capacity to run your own. Six: what is the smallest version that proves the case in 4 to 8 weeks, and what does measurable ROI look like by the end of that window? Seven: what is the kill-switch? If a chatbot’s hallucination rate moves above your threshold, or a workflow’s false-positive rate doubles, can you actually disable it inside an hour, or are you waiting on a vendor ticket?
The honest default for most UK SMEs in 2026 is workflow automation first, chatbot second, with the chatbot scoped tightly to a deflection band where it can fail without doing brand damage. Run the qualification framework in should you use AI for this before either, and build a one-page risk register for whichever path you pick.
