How to bring AI into your business
Most AI rollouts die between the demo and the workflow.
How we ship AI
- 1TestYour real tasks, run through Claude, GPT and Gemini. Scored before anything is built.
- 2BuildInside Gmail, Slack, your CRM or Sheets. Agents, MCP connections, retrieval on your documents.
- 3GuardLimits it cannot cross, a person on the exceptions, every action logged.
- 4MeasureHours and error rate before and after. If it does not pay, it comes out.
ClaudeGPTGeminiMCPAgentsRetrieval (RAG)Evalsn8n / ZapierYour CRM
Start from where the hours go, not from a tool. AI pays on high volume work with varied input, where a good draft saves most of the effort and a person still decides. On an unmapped process it only makes a demo.
By Shaheer Shaikh, technology and operations consultant · Updated October 3, 2026
Work it out in a minute
One task, four questions. Says which shape of AI fits, or that none of them do yet.
The plan
Do these six, in order
A time against each one and a way to tell it is finished.
List the ten tasks that eat the most hours
Mark each one as reading, writing, sorting or deciding.
2 hours · Done when all ten have an hours-a-month number
Pick one narrow use case
Text in, text out, with a person checking the output.
30 minutes · Done when you can say it in one sentence with an input and an output
Write the data rules before the pilot
What may leave your systems, what may not, and who approves an exception.
2 hours · Done when one page exists and an owner has signed it
Redesign the workflow around the tool
The model drafts and a person approves, rather than the tool bolted on the end.
half a day · Done when someone has walked the new sequence once
Name one owner and one number
Take the baseline before anything is switched on.
30 minutes · Done when the number has a before figure
Run four weeks, then decide
Compare against the baseline and either scale it or drop it.
4 weeks · Done when the number moved, or the task is off the list
The question most leadership teams are asking is not which model to use. It is whether any of this will still matter in two years, and what happens to the people currently doing the work. Both are fair questions, and neither is answered by a pilot that nobody adopts.
What follows treats AI the way you would treat any other operational change.
01.Map hours before models
Start with the same five-day hours log you would run before any process change. AI is a tool for specific tasks, and you cannot choose the task without knowing which ones consume the week. Look for reading, summarizing, drafting, extracting, classifying, and answering the same question repeatedly.
This also sets the baseline you will be judged against later.
02.Pick one narrow use case
One use case, narrow enough to describe in a sentence, with a clear before and after. Drafting first-pass replies to a common inquiry type. Pulling structured fields out of incoming documents. Summarizing calls into a standard format. Triaging a shared inbox.
Narrow is not timid.
03.Set the data rules first
Decide before anyone starts what may be put into which tool. Client data, contracts, financials, health information, anything under a confidentiality obligation. Write it down in plain language, one page, and say which tools are approved and whether their providers train on your inputs.
This is not a legal formality, it is what stops the project being shut down in month three.
Twenty minutes with a practitioner from our team, and you leave with a plan for your specific situation.
Talk to an expert04.Redesign the workflow around the tool
A tool dropped into an unchanged workflow gets used twice and abandoned. Decide explicitly where the model sits in the sequence, what it produces, who checks it, and what the person does with the time this frees.
Be equally explicit about the line the model does not cross.
05.Name an owner and a number
One person owns the use case and one number defines whether it worked. Hours returned, response time, error rate, throughput per person. Not a satisfaction survey and not a general sense that things feel faster.
Give it ninety days before judging.
06.Scale what the numbers approve
If the number moved, expand along the same shape. The same tasks in an adjacent team, or the next use case on the ranked list. Keep the pattern of one owner, one number, and a documented boundary, because that pattern is what made the first one work.
If the number did not move, say so plainly and stop.
Buying licenses before mapping the work.
Pilots that never leave the demo because no workflow changed.
Letting every team pick its own stack; sprawl eats the savings.
Questions we get
The ones that come up on almost every call.
Where do most companies get value from AI first?
Almost always in drafting, triage, extraction and summarizing.
Is our data safe in these tools?
It depends entirely on which tool and which plan.
How much should this cost to try properly?
A first use case is usually a matter of licenses and a few weeks of attention rather than a capital project.
What if the team is skeptical or worried about their jobs?
Take it seriously and be specific, because vague reassurance reads as evasion.
The first purchase is usually seats. ChatGPT Enterprise or custom AI says when that stops being enough. On what people may and may not put into a model, how to write an AI policy. Before the rollout, preparing for AI is the checklist that says whether you are ready to start one.
Most of this is doable in house.
Read next
The service behind this guide, the questions people ask, and the next thing worth reading.
Shaheer leads the work, with engineers, writers, filers and analysts behind him. C-suite operations for a San Francisco AI company, Six Sigma on the process side, Anthropic certified on the Model Context Protocol, ten years across eight industries. See what we have built