Best AI for a Growing Business: Standardise Processes Before You Automate

A growing business should standardise three things before adding AI: the steps of its busiest process written as a short SOP, the templates for the messages and documents it produces, and an intake form that captures the same details every time. The AI tool that then fits first is usually a chat assistant with project instructions, not a new platform.

The order matters because AI copies what it is given. If staff handle an enquiry five different ways, an assistant shown those five ways produces a sixth. If the intake form arrives half-filled, the assistant fills the gaps with plausible guesses. A written process turns "what we usually do" into something a tool can follow, and it makes the result checkable.

Follow me on Instagram@sagnikteaches

Why growth breaks the way you used to work

At three people, the process lives in the owner's head and gets fixed by conversation. At eight, each person has their own version, handovers lose detail and the owner becomes the checker for everything. That is the point where owners look for AI to take the load off, and where it tends to disappoint, because the tool is asked to do a job nobody has defined.

Connect on LinkedInSagnik Bhattacharya

The question is not "which AI is best", it is "which job is clear enough to hand over". Three signs that a job is ready: it happens at least weekly, it has an input you can describe (an email, a form, a spreadsheet), and a person other than the one who does it could say whether the result is right. Where those fail, write the process before touching a tool. The longer method for testing readiness is in how to tell if a process is ready to automate, and the cost of skipping it is spelled out in why automating a broken process backfires. The writing itself takes hours, not weeks, and it decides which tool fits afterwards.

Subscribe on YouTube@sagnikteaches

A worked example: a six-person HR consultancy and its enquiry reply

An illustrative firm: six people, three consultants, one office manager, one bookkeeper and an owner. New enquiries arrive by email, a web form and phone messages, and about 40 arrive a month. Each consultant answers in their own way. Replies take between 10 minutes and two days. Roughly one in four enquiries gets no reply because the consultant it landed with was on a client site.

The owner's first instinct is to buy an AI receptionist. The better first move costs nothing. Spend two hours writing down the current best practice:

1. Every enquiry is logged in one shared sheet within one working day, whatever channel it came by.
2. The office manager checks it against four questions: type of service wanted, number of employees, urgency, and whether it is a conflict with an existing client.
3. If the enquiry is a fit, a standard reply with two time slots goes out within four working hours. If not, a polite decline with one referral suggestion goes out.
4. The consultant who takes it confirms within one day and adds the call date to the sheet.

Four lines. The result is already worth something: the 25 per cent that never got a reply now has a named owner and a time limit. Only then does the question of AI arise. The logging step can use a form, the four questions can be a checklist in the form, and the reply is a template with blanks. An assistant can draft the filled-in reply from the form and the template in seconds, which the office manager reads and sends. A rough estimate: 40 enquiries a month at 12 minutes each is 8 hours; at 4 minutes with a drafted reply it is under 3. About 5 hours a month is saved, at no cost beyond a $20 plan, and because the steps are written, anyone covering for the office manager can do it. These figures are illustrative.

The lesson is the order: the checklist and template existed first; the AI shaved the drafting time. Had the firm started with the tool, it would have automated five ways of answering.

The four documents that make a process automatable

1. The SOP: numbered steps, owner, finish line

A standard operating procedure, or SOP, is a plain list of what happens, in order, and who does it. It does not need a policy-manual voice. Each step needs a verb, a person or role, and a clear end: "Office manager emails the standard reply and ticks the Replied column". The test of a good SOP is that a new starter could follow it without asking.

Include three things most SOPs leave out: the trigger (what starts the process), the exceptions (what to do when the input is odd) and the check (how someone knows it was done well). Exceptions are where AI goes wrong most, so write at least the top three.

2. The template: the words you would send every time

A template is a message or document with fixed text and marked blanks. The best way to write one is to take the three best past replies, strip the details and combine them. Put blanks in square brackets, such as [first name] and [service], and add a line in italics telling the user what to change. If you give an assistant this template, it follows it far more closely than it follows a description of tone.

3. The intake form: the same questions every time

Most reworking comes from missing information. An intake form, whether it is a web form, a spreadsheet row or a paper sheet, captures the same fields in the same order. Fewer than eight questions, each answerable in a sentence or a tick box, and one field for "anything else". The form is also the assistant's input: a model fed a complete form gives a better draft than a model fed a rambling email.

4. The decision rules: if this, then that

This is the document nobody writes, and the one AI most needs. Examples: "Quotes under $2,000 can be sent by the consultant; over that, the owner approves." "If the client is already late on an invoice, no new work is booked." "If the enquiry mentions a legal dispute, pass to the owner the same day." Write these as short sentences with a number or a condition. If you cannot write the rule, the decision is judgement, and a person keeps it.

Standardising with the help of AI, without handing it the pen

An assistant is a good editor of process documents and a poor author. You supply the facts, it supplies the structure. Three ways to use it while you write:

First, turn a voice memo into an SOP draft. Record yourself describing the job as you do it, transcribe it with your phone's dictation or a meeting tool, and paste it in:

Below is a transcript of me describing how we handle new client enquiries. Turn it into a numbered SOP with: trigger, steps (each with a role and a verb), exceptions, and a "done" check.
Rules: use only what is in the transcript. Where something is unclear, do not guess: list it as a question at the end for me to answer. Keep each step to one sentence.

The final rule is the one that protects you. An illustrative output has nine steps and four questions at the end, such as "Who decides when two consultants are both free?" and "What counts as urgent?" Those questions are the useful part, because they are the gaps in the process that were invisible when it lived in your head.

Second, ask for the checklist version. Paste the SOP and ask for a one-page checklist a person can tick, with no step longer than ten words. Print it or keep it pinned in the shared folder.

Third, test the SOP by asking the assistant to follow it on ten past cases. Paste the SOP and one anonymised past enquiry and ask what the process says to do, step by step. Where the answer is wrong or ambiguous, the SOP is unclear to a reader other than you. Fix the SOP, not the answer. Ten past cases is enough to reveal most of the holes.

Which AI tool fits first, once the process is written

For most small businesses at this stage the answer is a general chat assistant, set up with your written material. The main options in a project-style setup are ChatGPT Projects, Claude Projects and Gemini Gems, which Google is turning into skills (from 17 November 2026 on personal accounts and March 2027 on Workspace business accounts, per its help page). Each lets you store standing instructions and reference files so that every chat starts with your SOP and templates already loaded. Custom GPTs are scheduled to stop working on 11 December 2026, with OpenAI moving them to plugins, so do not build new ones.

The choice between them matters less than most comparisons suggest. Pick the one your team already uses for email and documents, or the one whose business plan your firm can afford. On OpenAI's pricing pages in November 2026, a ChatGPT Business Standard seat costs $25 a month billed monthly or $20 billed annually, from two users, and the individual Plus plan is $20 a month; Claude's Team seats are the same $25 or $20, for 2 to 150 people. Verify on the vendor's page before buying, since plans change. Business plans from OpenAI, Anthropic and Google do not train on your content by default, which is the point to check before any client information goes in.

The decision, by situation:

Note that the fifth row is the usual answer. Do not be embarrassed by it. A written process improves a business before any tool is added.

Standardising in other kinds of business

The method is the same everywhere; the documents change. A few short examples, all illustrative.

A mortgage broker with four advisers. The repeated job is the fact find: gathering income, outgoings and documents. The intake form lists the 18 questions every case needs, the checklist lists the documents by type of applicant, and the decision rule says when a case goes to the owner (for example, self-employed under two years of accounts). Once those exist, an assistant can read a completed fact find and list the missing documents. See the enquiry-to-fact-find workflow for mortgage brokers for that process end to end.

A financial adviser with a growing client list. The repeated job is the annual review. The SOP covers the pre-meeting pack, the meeting agenda and the follow-up letter, and the template has fixed sections with blanks for figures. The assistant drafts the follow-up letter from meeting notes, and the adviser edits the facts. Anything resembling advice stays with the adviser, which is written as a decision rule in the SOP.

An interior designer taking on more clients. The repeated job is the client brief. A one-page questionnaire captures budget, rooms, styles liked and disliked, and deadlines; a standard proposal template has three package options. Written once, they let an assistant draft a proposal from a returned questionnaire in minutes.

A common thread: in each case the written parts took one to three hours, and the AI step was small. Most of the gain, fewer reworks and a faster reply, came from the writing.

A filled-in intake form and the draft it produces

To see how the pieces fit, here is the HR consultancy's form after one enquiry, again illustrative. Service wanted: handbook review. Employees: 34. Urgency: needs it before a new site opens in six weeks. Existing client: no. Other: "worried about our probation wording".

Given that form, the standard reply template and the decision rules, an assistant returns a draft such as: "Hi [first name], thanks for getting in touch about your handbook. With 34 people and a six-week deadline, the first step is a 30-minute call to look at the probation wording and agree scope. I can offer Tuesday at 10:00 or Wednesday at 14:00. [your name]". What the office manager fixes: she checks the two slots are really free, and she removes any promise about scope that the form did not support. Two edits, thirty seconds. Compare that with the same enquiry arriving as a rambling email with no deadline mentioned: the draft would have asked three questions instead of offering two slots, and the exchange would take a day longer. The form did the heavy lifting; the assistant only wrote it up.

Where standardising goes wrong

The first failure is the document nobody follows. An SOP written by the owner over a weekend and emailed to the team gets read once. Involve the people who do the job: ask each to record or describe one real case, and write the SOP from what they say. Put it where they already work, not in a folder that nobody opens.

The second is the over-written process. A 12-page SOP for a five-minute task will not be used. A useful limit: one page, ten steps at most. If it does not fit, split the process in two.

The third is writing the process as you wish it worked, not as it does. An assistant will happily follow the wished-for version and produce drafts that staff cannot send because they depend on a step nobody takes. Observe the real process for a week first, then write the improved one.

The fourth is the frozen document. Processes change. Put a review date on each SOP, assign an owner for each, and ask after each real problem: "should the SOP have said something different?" If yes, change it that day.

A two-week plan for the first three processes

Days 1 to 2: list the repeated jobs. Ask each person to name the three tasks they repeat most and the one they dread. You will find the same two or three names repeated; those are your candidates.

Days 3 to 5: choose one and write its SOP from a real case. Use the voice-memo method above. Review the draft with the person who does the job, and have a second person follow it once without help. Fix every place they stop.

Days 6 to 8: write the template and the intake form for that process. Test the template by sending three real replies from it.

Days 9 to 12: add the assistant. Create a project, paste the SOP, template and decision rules as instructions, and run ten past cases. Count the drafts that could be sent with one small edit; if fewer than seven in ten, fix the instructions, not the model.

Days 13 to 14: measure. Time the process before and after on five jobs, and note errors. Then decide whether to repeat the exercise on the second process or let the first one run for a month. A scoring page for choosing the second job is in the one-page AI use case template, and the "what next" choice is covered in what a small business should automate first.

Reading the result: did the process or the tool help?

It is easy to credit the tool for what the writing achieved. Separate them: measure the process after you write it and before you add AI, then again after. In the HR example, the first measure would be the reply rate and the time to reply. If the rate moved from 75 per cent to 100 per cent after the SOP and the time stayed at 12 minutes, the writing did the first job, and the assistant did the second by cutting time to four minutes. Record both. When a tool vendor later offers an upgrade, you will know what a tool actually gives you for the money.

If you would rather map this with someone, that is what my AI implementation consultation covers: choosing the three processes, writing them with you and matching each to a tool you may already have.

Questions from owners who are growing faster than their processes

How many processes should I write down before starting with AI?

Three is enough to begin: the one that wastes the most staff time, the one that most often produces complaints, and the one a new starter finds hardest. Write each as numbered steps with who does what and what a finished job looks like. Add more once the first three run smoothly.

Can AI write my SOPs for me?

It can draft them from your rough notes, voice memo transcripts or an observed walkthrough, and that saves hours. It cannot know what you actually do. Someone who does the job must read each draft against reality and correct it, because an invented step in an SOP is worse than a missing one.

Do I need a paid AI tool to standardise processes?

No. The standardising work needs a document editor and a shared folder. A free chat assistant can help tidy drafts. Pay only when the process is stable and a paid plan adds something you need, such as team sharing, longer files or a workspace where staff can use the shared instructions.

What if my team does the same job three different ways?

Pick the version that produces the fewest corrections, not the one the longest-serving person likes. Ask each person to record one real example, compare the results, and write the SOP from the best. Tell the team why that version was chosen so the rule is not read as a criticism.

When is a process too messy to automate at all?

When two people given the same input would produce different outputs and neither could say why. That means the decision rule is missing, not that AI is the wrong tool. Write the rule, test it on ten past cases, and then automate. If no rule can be written, keep a person in the loop.

Further reads

• How to Document Your Processes Before Adding AI — A fuller method for documenting a process before AI.
• How to Build the FAQ Your AI Chatbot Needs Before Launch — Turn your written answers into an FAQ an assistant can use.
• AI Tool Approval Process: How Staff Request a New AI Tool — Let staff request new AI tools through one route.
• A Simple AI Risk Register for Small Businesses (With Template) — Record what could go wrong as you add AI to a process.
• Quote to Cash: Connect Quotes, Invoices and Payments With AI — A full process from quote to payment, with AI steps.
• AI Tools and AI Development: The Complete 2026 Guide — the AI hub, including every tutorial in the AI-for-business series.

Sources: OpenAI ChatGPT pricing pages and custom GPT retirement FAQ, Claude pricing page, Google help page on the move from Gems to skills, and the three vendors' training-data pages (checked 5 Nov 2026). The SOP examples and numbers are illustrative.