A year of using AI leaves most small businesses with a mixed pile: a few habits that quietly helped, a few mistakes that taught more than any tutorial, and a longer list of shiny ideas that never earned a place in the week.
December is a good time to sort that pile. Not to declare a transformation. To write down what helped, what hurt trust, and what 2027 should do differently with people still making the calls that matter.
This post shares AI lessons from 2026 using an invented example business. It is illustrative, not a client story. Features of the WeKinnect Marketing Platform depend on the setup WeKinnect activates for a client, so treat every tool description as something to confirm for your account.
Iris runs a small pottery studio with two wheels, a slab table, a kiln room, and a tiny retail shelf of finished mugs and bowls. She teaches beginner throwing classes, offers glaze workshops, takes custom dinnerware orders, and sells a limited number of finished pieces. A kiln tech comes in twice a week. A front-desk helper covers class check-in on busy evenings.
In 2026 she turned on a few AI-assisted tools through her marketing setup:
Some of it helped. Some of it embarrassed her. All of it taught something.
Summer is for making pots. December is for noticing which tools earned a place beside the wheel. If Iris waits until a busy March to review AI, she will only notice the tool when it fails in front of a customer. A year-end look lets her change rules while the studio is relatively calm.
The tone of this review is practical. No transformation story. No claim that AI reinvented the studio. Only lessons that change next week’s habits.
The clearest win was boring. People asked when classes started, whether beginners were welcome, whether the studio provided clay, and how to reserve a wheel. The chat and the drafted replies handled those well when the knowledge base was current.
The win was not cleverness. It was coverage. Iris could teach a class without missing every inquiry that arrived during demos. The human step remained: she reviewed transcripts weekly and fixed answers that sounded off.
Keep: AI on repeated, factual questions with a short approved answer list.
Fix: any answer that drifts when the schedule changes. Stale knowledge is worse than no chat.
When a drafted reply once told a customer the wrong firing day, Iris corrected it by hand and apologized. Most customers were fine. When an automated message repeated a similar mistake before she caught it, the trust cost was higher. People expect tools to be consistent. When they are wrong, the wrongness feels colder.
That pattern matches what many owners notice without a lab: people forgive a corrected human mistake more readily than a repeated automated one. Iris’s practice follows from it: AI drafts, humans approve anything that names a date, a price range, a safety rule, or a promise.
Keep: human approval on anything that can be wrong in public.
Leave alone: fully automatic replies about kiln safety, custom quotes, or anything medical or legal. Pottery has fewer of those than some fields, but glaze safety and workshop age limits still need a person.
A large study of generative AI in customer support, summarized by the National Bureau of Economic Research as Generative AI at Work, found productivity gains that were larger for newer and less experienced agents than for the most experienced ones, along with improvements in customer sentiment in that setting. The paper is about a specific workplace and tool. It is not a promise about a pottery studio.
What Iris takes from it is modest and practical: AI drafting helped her newest front-desk helper answer routine questions faster, while Iris herself still rewrote almost every custom-order reply. The tool was a training wheel for repeated work, not a replacement for judgment on commissions.
Keep: using AI drafts to help newer staff learn the studio’s plain answers.
Fix: any setup where a new helper can send without review.
Every bad AI answer Iris saw traced back to missing or outdated information. Class calendars not uploaded. A glaze workshop moved without updating the site. A “we provide tools” line that was true for throwing class and false for some workshops.
Vendor guidance for AI employees usually includes the same parts: a written role, channels, capabilities switched on or off, and knowledge sources such as the website, uploaded documents, and basic business facts. An explanation on a response can show which sources were used. Conversation logs show questions the AI could not answer well.
Iris’s year-end rule: if it is not written down in the knowledge base, the AI should say it does not know and offer a human follow-up.
Keep: a monthly knowledge-base check on the same day as the listing hours check.
Fix: long, vague studio descriptions that invite the AI to invent a friendly tone without facts.
A chat that cannot reach a person becomes a wall. Iris’s best setup change was a clear handoff: if the question is about a custom order, a complaint, a class scholarship, or anything the knowledge base does not cover, the tool offers to take a name and have the studio reply.
Depending on platform edition, conversations can be assigned to staff, create follow-up tasks, and notify by email or a daily digest. Iris uses the simple version: a named inbox she checks between classes.
The lesson: AI that captures and routes is more valuable than AI that performs certainty.
Customers did not mind that a first reply might be assisted. They minded when a message pretended to be Iris in a personal way and then failed a simple fact. Her practice by year-end:
Keep: sounding like a studio, not like a motivational poster.
Leave alone: AI-written brand stories. Iris writes those.
There are places Iris will not put AI in 2027:
Lead scoring, where it exists, is configured on the agency side for some subscriptions and is not confirmed for every client account. Iris does not promise it, and she does not want a silent score deciding who gets a reply first. People still own that call.
Iris tried AI-drafted social captions for finished mugs. The drafts were fine and forgettable. When she rewrote them with one true detail — the glaze that ran, the student who centered for the first time — the posts sounded like the studio.
Keep: AI for a first skeleton when she is tired.
Fix: posting the skeleton without the true detail.
Leave alone: letting a tool invent student stories. Permission and truth still matter.
Calendar booking into a connected Google or Outlook calendar can help when it is enabled. Iris allows it for studio tours and beginner class waitlist calls, with buffers so it cannot double-book kiln days. She does not allow it to book custom-commission consults without a form she reads first.
The lesson: automation may fill a slot. Only a person knows whether that slot should exist that week.
An illustrative review Iris can repeat:
When the performance report is connected, Iris can see messages received versus messages sent by AI and by the team, depending on setup. That split is useful. It shows whether the tool is carrying routine load or going quiet.
She does not turn those counts into a boast. She asks:
If the report cannot answer those questions, she falls back to the conversation logs. The logs are the ground truth.
Depending on setup, the platform can support an AI workforce with named roles, channels, knowledge sources, simulated tests, conversation logs, and handoff to staff. Appointments can be booked into a connected calendar when that is enabled. Reports can show messages received versus messages sent by AI and by the team.
None of that removes Iris’s review habits. The platform makes the review possible in one place. She still decides what “good” means for her studio.
Organizations looking for a structured way to think about AI risk can look at the NIST AI Risk Management Framework Playbook, which offers voluntary suggested actions aligned to Govern, Map, Measure, and Manage functions. It is not a checklist Iris must complete, and it is not a certification. It is a reminder that AI work has governance, mapping of context, measurement, and ongoing management — the same ideas as her weekly log review, written in public-sector language.
For a pottery studio, that translates to: know who owns the tool, know what data it uses, know how you will notice mistakes, and know how you will change it.
Before Iris locks 2027 rules, she confirms:
Those questions prevent a January surprise.
May draft: class times, what to bring, whether beginners are welcome, parking notes, links to register.
Needs approval before send: anything with a date change, fee information, waitlists, workshop age limits.
Human only: custom quotes, complaints, kiln/safety calls, scholarships, partner or wholesale inquiries.
Weekly: read logs; fix knowledge base; note new repeated questions.
Monthly: confirm hours and class calendar in the business profile and on the site.
An invented volunteer-run children’s museum used a simple chat tool on its donation and visit pages. It is illustrative, not a real museum.
What worked: hours, exhibit closures, and party-room rules answered after closing time.
What failed: a draft that guessed at accessibility details the volunteers had not written down.
What they changed: only published facts go in the knowledge base; anything about access, medical needs, or donations routes to a person.
Iris’s lesson from their year matches hers: AI is allowed to be helpful. It is not allowed to be confident about facts nobody confirmed.
A few 2026 moments stayed with Iris:
None of these were dramatic. Each one spent a little trust. The fix was always the same: update the source, tighten approval, slow the send.
Calling AI a workforce is useful only if it means roles, handoffs, and review — the same words you would use for a person. It is not useful if it means “the studio runs itself.”
Iris’s AI chat is closer to a trained apprentice who may answer the phone about hours and must fetch Iris for everything else. That metaphor keeps her expectations sane.
She will not:
She will:
Iris’s short list:
Some owners reading this will think they are late to AI. Iris felt that in the spring. The useful response was not to turn on every feature. It was to pick one repeated question type, write the answer once, and let a tool help deliver it with supervision.
Being behind on features is not the same as being behind on judgment. Judgment is still the scarce part.
Do not claim the studio runs on autopilot. Do not claim AI increased sales by a number you did not measure. Do not claim a tool “knows” customers. Say what it drafts, who approves, and where it stops.
Iris’s kiln tech does not need to run the chat tool. Her front-desk helper does. Training took one hour:
The helper’s confidence went up when she could modify drafts, not when she was told the tool was smart. Control mattered more than awe.
Iris’s test for whether a feature belongs: turn it off for a quiet month and see what she misses. If she only misses the feeling of being modern, it can stay off. If she misses fewer dropped class questions, it belongs, with rules.
She did that with auto social replies, which she never fully launched. She did not do that with class-hours chat, which she would miss immediately.
The test is humble and effective. It prevents tools from remaining on by inertia.
A year-end AI post can sound like a confession of failure. Iris refuses that tone with her team. 2026 was a year of experiments. Some paid for themselves in time. Some taught boundaries. Both are valuable. The point of sharing lessons is to make 2027 kinder to the people who answer the phone, not to scold last year’s owners.
Iris writes one sentence on a card by the front desk:
Tools may draft the ordinary answers; people still own every promise.
That sentence is the 2026 lesson. It is also the 2027 rule.
The best AI lesson from 2026 for Iris was not technical. It was relational. Tools that helped her answer faster were welcome. Tools that spoke for her without her judgment were not.
If you’d like help reviewing your AI setup from 2026 and writing clear rules for 2027, with people still holding the decisions that matter, start a conversation with WeKinnect.
Photo by Anne Nygård on Unsplash
Written By
Maren Hale focuses on brand positioning at WeKinnect Global Branding Agency. Guided by the core value of Service before Self, she writes for entrepreneurs and family-owned legacy businesses. Maren writes on Business Listings, Google Business Profile, SEO, GEO (being found by AI tools like ChatGPT, Gemini, and Grok), and Community. WeKinnect serves businesses across restaurants, arts, travel and tourism, banking, home services, nonprofits, and the pet industry. Maren brings that same spirit of service to every piece she writes.
Darlene Gagnon
16 Oct 2026
