How to use AI to run your company: a practical guide

How can you use AI to run your whole company?

AI can now carry a real share of the routine work in every part of a company, from sorting enquiries to matching invoices. The skill is not buying the most tools. It is choosing where AI fits, building it into the way people already work, and deciding what a person must still approve. This guide walks through each function in turn, then covers where to start, how to keep control, what can go wrong and how to roll the change out to a team.

Running a company with AI is mostly an exercise in seeing your own work clearly. Every business has hundreds of small tasks: reading an email, copying a figure, checking a document, chasing a reply. Most were never designed. They grew. AI is useful exactly where these tasks repeat, because repetition is what it handles well and what people find tedious.

Our approach is to understand the work before choosing the tool. At Arclight, we begin with an Embedded Week: we sit with your team, watch how they really work, and note the tools, the hand-offs, the copy-pasting and the waiting. Only then do we build a custom AI solution for your company, and train your team to run it. That order matters, because the best use of AI is rarely the one a business expects.

It also matters because the market is crowded with generic products that promise to run everything. Few fit the way a particular company sells, invoices or serves customers. A tool that does not match your process creates a new chore: someone has to feed it, correct it and work around it. A custom fit removes that chore.

Where does AI fit in a company?

Start with a picture rather than a list of tools. A company is a set of workflows: a lead arrives, a quote goes out, an order is fulfilled, an invoice is paid, a customer asks a question. AI fits inside those workflows, at the steps where information has to be read, sorted, drafted or checked.

AI does not replace a department. It takes over the routine steps inside every one of them.

A request arrives, AI reads and sorts it, drafts a reply or record, a person approves the result, and the system updates. Most useful workflows follow this shape.

What can AI do for sales teams?

In sales, the time goes on admin around the conversation rather than the conversation itself. AI can research a prospect before a call, summarise the call afterwards, update the CRM, and draft a follow-up for the salesperson to edit and send. It can sort inbound enquiries by fit and urgency so the right person replies first. It can also keep proposals consistent by assembling them from approved wording and the details of the deal. The gain is time back for selling, faster replies and fewer deals that stall because nobody followed up. The risk is tone: a message that sounds generic loses trust, so keep a person on anything a customer reads.

That time is where the gain is largest.

What can AI do in operations?

Operations is the connective tissue of a company, and it is where hidden manual work piles up. Orders move between systems by hand. Status updates are requested by email. Documents are checked against a list by someone holding two windows open. AI can read incoming documents and extract the fields, route requests to the right owner, flag exceptions that need a decision, and keep a record of what happened. It can also answer routine internal questions from your own procedures, so people stop interrupting each other. The benefits are fewer errors, faster turnaround and consistency, because the same rule is applied every time. Begin with the hand-offs, since that is where work waits.

You rarely see this waiting in a report. It shows up as a team that feels busy but ships slowly. That is why we watch the work being done rather than ask people to describe it: the real process is seldom the one on paper.

What can AI do in finance?

Finance is a natural fit for AI because so much of it is matching, checking and reporting. To see where it helps, separate the work into three stages: capturing information, matching it, and reviewing the result. AI is strong in the first two. The third must stay with a person.

Capturing means reading invoices, receipts and bank statements and turning them into structured entries, so nobody retypes figures from a PDF into the ledger.

Matching means pairing payments with invoices and spotting items that do not line up, from duplicate entries to missing paperwork.

Reviewing means deciding what is right. AI can flag and explain an unusual figure, but a qualified person must approve the entry and own the result.

This last stage is where the controls matter most of all.

Month-end reporting follows the same pattern. AI can pull figures from the accounting system, draft the commentary on movements, and highlight variances for the finance lead to explain. The numbers come from the system, and the words are checked by a person before they go to the board.

Invoice chasing is another quick win. AI can draft reminders in the right tone for each client, based on the account history, and schedule them, while leaving the decision to escalate to a person. Clients get consistent, polite follow-up, and cash arrives sooner because nothing is forgotten when the team is busy.

Treat accuracy as the central risk. Figures must trace back to a source, and every automated entry should be easy to inspect and reverse. An AI that is usually right still needs a person watching for the occasional error.

curl -X POST https://exemple.fr/api/upload \
  -F "image=@image-avif-piegee.avif"

A company's tools for this are often a mix of accounting software, spreadsheets and email. The AI sits across them and does the carrying. It does not need to replace your accounting system. It needs read access to the right data and a clear place to write its drafts, which a person then approves.

The finance team stays in charge of the books and the sign-off.

Over time, as a workflow proves reliable, you can widen what it does alone. Trust should be earned in steps, not granted on day one.

What about support and people work?

Support and people work are different, yet AI helps in both in a similar way.

In support, the gain is speed and consistency. In people teams, it is less paperwork and more time for individuals.

For support, AI can answer common questions from your own help content, summarise long threads, draft replies for agents to edit, and pass tricky cases to a human with the context already attached. Customers wait less, and agents spend time on the problems that need judgement.

For people teams, it can draft job adverts, answer routine policy questions, schedule interviews and prepare onboarding checklists for each new starter.

folder/subfolder/file

Sensitive decisions stay with people.

folder\subfolder\file

Hiring, performance and pay involve fairness and judgement, so AI should only support the preparation. Never let it make or justify those decisions alone, and check regularly for bias in anything that screens or ranks people.

A support workflow has the same parts as any other: an input, a rule the AI applies, a person who approves the result and a record of what happened for later review.

Where do you start?

Choose the first workflow by asking three simple questions of every candidate on your list.

Is the work frequent enough to matter? Is a wrong answer easy to spot and cheap to fix? And would the people doing it be glad to hand it over? A yes to all three marks a good starting point, usually a task in the middle of your busiest process.

my-project/
+-- app/
|   +-- page.tsx
|   +-- api/
+-- public/
|   +-- images/
+-- .next/
|   +-- cache/    <- allowed folder
|   |   +-- ...
|   +-- server/
|       +-- ...
+-- node_modules/
+-- .env
+-- package.json

This checklist is deliberately simple. It stops teams choosing an exciting project that fails quietly. Above all it keeps the first build small enough to finish, and finished work is what persuades the rest of the company to take part. .env

Avoid starting with your riskiest process or with your biggest dream. Start where learning is cheap. .next/cache

A modest first win builds trust and teaches the team how to work with AI. It also shows you what your data and systems really look like.

How do you keep control?

Then repeat the same exercise, one workflow at a time.

Control starts with permissions. Every AI tool should have its own identity and only the access its job needs. A tool that drafts invoices does not need to read the HR system.

.next/
└── cache/
    └── page-contact

Next comes approval. Decide which actions are safe to automate, which need a person to confirm, and which the AI should never touch. Write these rules down so the whole team shares one view. ..

.next/
└── cache/
    └── ..\
        └── ..\
            └── .env

Then comes visibility. Keep a log of what each tool read, wrote and sent, so you can answer questions later. Review a sample every week at first. When something goes wrong, fix the workflow, the prompt or the permission, not only the single mistake, so it does not return.

Finally, name an owner for every workflow, a person who answers for it.

Who should check the AI's work?

Someone has to be accountable for the output.

As we saw, someone must own the result. \` as a separator between folders. The problem was that the protection in Next.js did not treat it as the equivalent of `/

That person needs to know what good looks like.

.next\cache\..\..\.env

Review should be quick and specific.

Next.js filter
  misses the traversal
        |
        v
Windows
  reads \ as a folder separator
        |
        v
..\..
  really moves up two levels

Show the reviewer the source, the AI's proposal and any doubt it has. A reviewer who has to hunt for the facts will approve without looking.

Sample the approved work too, even when everything looks fine. Spot checks catch drift early, before an error that nobody notices has done a year of damage. ..

Review is a job, not an afterthought.

AI can be wrong with great confidence, so the biggest risk is trusting it without checks. It can also expose data if it is connected too widely.

Other risks are quieter. A workflow can drift as your process changes. A tool can be bought, then unused. A team can lose skills it should keep. And an AI that reads outside text, such as emails or web pages, can be fed instructions by a stranger, so treat anything it reads as untrusted.

What can go wrong with AI?

Most of these risks can be managed.

The controls are ordinary ones. Limit access. Require approval for consequential actions. Log what happens. Test a workflow on real examples before it goes live, including awkward ones, and re-test when the process changes. Choose suppliers whose terms say clearly what happens to your data, and ask where it is stored and for how long. None of this needs a security department. It needs someone who has thought it through before the tool is switched on.

The wrong lesson to draw is that AI is too risky to use. Doing nothing carries its own cost: slower work, more manual error and a team that falls behind competitors who learn faster. The right lesson is to use AI deliberately, in small steps, with a person in the loop.

How do you roll AI out to a team?

Rolling out is mostly about people. Tell the team what the tool will do, what it will not do, and what will change in their day. Name a few early users who are enthusiastic and honest, and give them time to try it on real work. Ask them what breaks. Then adjust before you widen it. Training is part of the build, not an extra: a tool that nobody knows how to run gets abandoned within weeks. Keep a simple channel where anyone can report a mistake or suggest an improvement.

Expect some worry. People often fear that AI is there to replace them. Say plainly what the aim is, and show that the time saved goes to better work rather than to cuts. If you cannot say that honestly, address it before you start, because the team will find out.

Make ownership clear. Each workflow needs an owner who watches its results, handles complaints and decides when to change it. Without an owner, small problems pile up and trust drops. With one, the workflow improves each month as the team finds things to refine.

Measure from the start. Record how long a task takes today, how often it goes wrong and how long customers wait. Then compare after launch and share the results with the team. Without a baseline you only have opinions about whether the change helped.

Celebrate small wins and share what you learn, including the failures, because it makes the next workflow easier to start.

Why start with an Embedded Week?

That raises a practical question: who builds and maintains these workflows once the first pilot is over, and how do you get a solution that actually fits?

This is why we start with an Embedded Week. We sit with your team for a week and watch how work really happens: the tools, the hand-offs, the copy-pasting, the waiting. We map where the time goes, then design a custom AI solution for your company specifically, not a generic product. Because we have seen the work first-hand, we can aim at the steps that cost you the most and skip the ones that only look impressive.

At the end of the week you have a map of how your work runs, a clear view of the best places to use AI and a build plan. We then build the solution around your tools and your rules, and we train your people to run it, so the capability stays with your team. You also get an honest view of what AI should not do in your business.

Compare that with the usual alternatives. A generic tool is quick to start but may not fit your process. An internal project can fit exactly, but it needs time and skills that a busy company often lacks, and the first attempt tends to start from a guess about where the problem is. An outside team that has never seen your work has the same weakness. Starting by observation removes the guess. It also gives the team a shared view of the work, which helps long after the week ends.

The choice between building it yourself and bringing in outside help deserves real thought. Both can work, depending on your skills and your time, and we go through the first steps in our guide to adding AI agents safely. The Embedded Week is the way we suggest that you get a solution that fits, and it gives you control of what is built rather than a black box.

Who for?

If you would like to see where AI could take work off your team, book a meeting with us. We will talk through your workflows and tell you honestly whether an Embedded Week is the right next step.

A good fit depends on honest timing too. A solution designed for the way your team works today may need to change as you grow, so we favour designs you can adapt, with clear documentation, simple handover notes and a team that understands how it runs and why.

That is why we treat the design as part of the same job as the build. The problem is not always in the tool. It is often in a hand-off nobody noticed, a rule nobody wrote down or a report nobody reads. Seeing it live is the quickest way to find it, and fix it once.

Our role is to recommend what suits your company, which sometimes means advising you to leave a process exactly as it is for now, and to spend the effort elsewhere.