AI for the work that eats your team's time.
Your team spends hours on data entry and cleanup, reporting and the gaps between systems. I build AI into those jobs so they run reliably and a person checks what matters.
Where AI pays off
The best candidates are jobs that are tedious for people, have a clear right answer, and happen often. The model does the reading and matching. Code does the validation and the writes. A person checks the cases that matter.
- Data processingOrders, invoices, product feeds, spreadsheets and PDFs that arrive in the wrong shape. The model extracts and normalises them and flags what it is unsure of. Clean records go into your system through a validated path.
- CRM, ERP and EDI integrationsConnections where the field mapping is messy and the exceptions pile up. AI handles the fuzzy matching and the exception notes. Code handles every write, with a record of what happened.
- Dashboards and analyticsAsk a question in plain English and get a chart. The model writes the query, code computes the numbers, and every figure in the answer traces back to the data.
- Classification and triageSupport requests, leads, documents, products. The model proposes a category with a confidence score, and a person approves the ones that matter.
- Content from your dataAlt text, product descriptions, summaries and first drafts generated from records you already hold, reviewed before they ship.
- Internal tools and agentsSmall tools that take a recurring job off someone's plate, with a defined scope, a log of what they did, and an off switch.
How I keep it reliable
- Structured outputs validated against a schema before anything downstream sees them
- Numbers computed by code; the model writes the query and explains the result, and any figure that cannot be traced is flagged
- Confidence thresholds, with low-confidence cases routed to a review queue instead of guessed
- An evaluation set with known answers, run before launch and after every model change
- Untrusted input, such as third-party text or uploaded documents, wrapped so it cannot steer the model
- A model chosen per task, cost logged per feature, and a way to switch the feature off
- PHI and PII mapped before any record reaches a model. I have been a HIPAA compliance officer, so the data-handling decisions get made at design time and written down
What to avoid
- Models that change production systems with nobody approving
- A chatbot with no defined scope or knowledge boundary
- Sending sensitive data to a service without a plan for it
- Features that cannot be switched off if they misbehave
The approval-gate pattern
For any AI feature that can change something real, the model proposes, a person approves, and the system verifies the result. The model has no direct write path. This is the pattern ZoneSteward uses, where every change is approved on the server and checked against the live system afterward.
Work I have done
- ZoneSteward: the model writes Cloudflare analytics queries, the server computes the figures, and every number in an answer is traced; alert rules drafted from a sentence and dry-run before they are armed; AI incident investigation with a suggested fix behind an approval card
- CookieSteward: unknown trackers classified by AI with structured output and a confidence score, hardened against prompt injection, with a person approving each verdict
- An alt-text plugin on the Claude API, deployed across an agency's whole client portfolio
- AI-assisted development adopted across a web team, with standards and review to match
- Years of the plumbing these features sit on: ERP and accounting sync, TrueCommerce EDI, point-of-sale and CRM integrations for stores and B2B businesses
Cost and data handling
AI features have running costs and data implications, and both should be decided up front. I scope usage costs, choose which model fits each task, decide what data leaves your systems, and make sure there is a way to monitor and limit all of it.
A fixed-scope AI audit
If you are not sure where AI would help, I can run a short audit. I look at your processes, data and systems, identify the jobs where AI would pay off, describe how each would work with a person in control, and estimate effort and risk. You leave with a plan whether or not we build it together.
Related
Can AI really handle our order and product data?
Yes, for the parts that are pattern work: reading, matching, normalising and flagging. The write into your ERP or CRM is ordinary validated code, and anything below a confidence threshold goes to a person.
Which models do you use?
Mostly Claude through its API, and I pick the model per task. The right choice depends on quality needed, speed, cost and data requirements.
Will it make things up?
Models can be wrong. That is why numbers are computed by code, outputs are checked against a schema, and a person sits in the approval step for anything consequential.
Can you add AI to my existing WordPress site or store?
Yes, as a plugin or a connected service, with the same approval pattern and clear limits on what it may do.
What happens to my data?
We decide it together before building: what is sent, to which provider, under what terms, and what is stored. Nothing is sent that you have not agreed to.