This is the second post in a series that works through one industry at a time, looking at how they can integrate AI. This week I’m looking at professional services firms.

When I write about manufacturing, I’m writing those articles based on over thirty years of experience in that sector. Professional services are different for me, and I want to be upfront about that. I now run a professional services firm, so I write the proposals, scope the engagements, and dig through old files looking for the one I know I wrote previously. I’ve also built knowledge tools for professional services firms, so I have first-hand experience in seeing how this plays out inside real practices.

I want to start by reiterating my standard caveat. I (nor any other vendor) can’t tell you from the outside which solutions fit your firm — that requires time working together. However, I can point at general areas where my experience says AI tends to earn its place in a firm that bills for its expertise.

Two things make this environment different from a shop floor. The first is that a lot of the work AI can touch in a firm is billable, so that may change where you point it first. The second is that a professional services firm holds information you must be hyper careful about feeding into an AI tool. Both of those are important, and I’ll come back to them.

The Work That Eats the Week

Anyone who runs a firm knows the parts of the week that don’t directly result in billable outcomes — research, chasing engagements, intake and triage, answering the phone, staying on top of the firm’s activities.

Much of that work is mundane and time consuming, and a lot of it is work you can’t bill for. That makes it worth focusing on as you begin to identify the right places to deploy AI.

Where I Would Point It

Here are five places, roughly in the order I would look at them.

1. Putting your firm’s knowledge within reach

A firm sits on a deep store of its own material: precedents, prior engagements, working papers, internal memos, and the standards and regulations it relies on. The answer to most internal questions is in there, but it can be slow to find. A constrained retrieval tool (the category our own Docora product sits in) answers only from the documents you approve, shows you where each answer came from, and tells you when it does not know rather than inventing something. It is quick to stand up against your own document set, which makes it a sensible first step. It matters more here than almost anywhere, because a wrong answer with a fake citation attached is a liability. The same engine can also face outward as a public assistant on your website.

2. Drafting the things you write over and over

Most firms rebuild a lot of their documents from templates, and that is where AI can shine: A custom app pulls from your intake and client records, applies your pricing and scoping rules, and assembles a branded, customized draft for a person to review and send. It’s important to note that this is a custom AI-enabled logic engine that encodes how your firm prices and scopes work, not a document editor. The same approach handles the documents around those standard templates: client letters, memos, routine agreements, financial-statement narratives, and the standard sections of recurring filings, all drafted from your own templates and prior work, then edited and signed by a qualified person.

There are a couple of cautionary notes necessary here. This kind of tool is a drafting accelerator, never an autonomous author that works without human signoff. And because a lot of this work is billable, you need to decide whether the time you save becomes more capacity or a change to how you price.

3. Intake and triage

This area includes new-matter intake, conflict and similar checks, summarizing what an incoming client actually needs, and routing it to the right person. This work is high in volume, mostly non-billable, and a strong early win for the same reason document retrieval is.

4. The phone and the front desk

Firms still field a steady stream of bounded, repetitive calls: scheduling, status, basic intake, and after-hours coverage. An AI voice system handles those well when the integration underneath is done properly. I covered this one in detail recently, so I’ll point you back to that piece rather than repeat it.

5. The ambitious one, an assistant that watches the whole practice

This is the biggest project on the list with a significant scope of work. Picture an assistant that sits across the firm’s email, calendar, sent items, and billing and time records, and each morning gives the managing partner a short, ranked read on what needs attention: client emails that have waited too long for a reply, commitments made in writing that are coming due, matters where the work is outrunning the budget, time worked but not yet billed, invoices going unpaid, and clients who have gone quiet. It surfaces and suggests areas of the business that might need attention, but a person decides and acts. You build an app like this in stages, starting with a read-only view and adding more functionality only once it has earned trust.

You can extend the same idea to the conversations on your Slack or Teams channels, so the partner gets a real pulse on what is happening across the firm without reading every thread. Be aware that this is where the effort and the sensitivity both climb — giving an autonomous agent that much visibility is not a decision to make lightly. A full picture means the assistant has to see all staff email and chat. Going this far is not for everyone, and it requires full disclosure to your team.

Making Sure the Draft Is Right

The moment you move from retrieval to drafting, the risk changes. A retrieval tool that only answers from your own documents can fail safe, but a drafting tool generates new language, and that’s exactly where AI can produce something that reads well but is simply wrong. That’s not a minor concern in professional work: there have already been well-publicized cases in Canada of lawyers filing documents that cited cases the AI had invented, and a fabricated citation is the kind of public failure that damages the trust a firm is built on. The technique I rely on most is having a second AI model check the first one’s work, tasked with finding every citation, figure, or claim the source documents don’t support, backed by plainer non-AI checks and a simple principle: match the rigour to the consequence.

Where Your Client Files Can Go

Everything above assumes the AI can see your documents, and in a firm those documents are the most confidential information you hold. So the first real question is not which tool, but where your client files are allowed to go.

Putting confidential client material into a public, consumer-grade AI tool is hard to square with a firm’s obligations. The personal information in those files is governed by privacy law, and on top of that lawyers and accountants carry professional duties of confidentiality, with privilege a live concern for legal work. The concern here is less about a single rule than about a stack of obligations.

In this area, consumer or free tools (where your inputs are almost always retained or used to improve the model) are off the table for confidential client work. You either need a business arrangement that contractually guarantees your data is not used for training, is held with appropriate residency, and leaves an audit trail, or, for the most sensitive files, a private or self-hosted model where the data never leaves an environment you control. For a small firm, a private deployment in a Canadian region is usually the sensible middle ground. It costs more and takes more setup than signing up for a consumer tool, but that is simply the price of handling client data correctly.

Note that I am not giving you legal advice here and the details differ by profession and province, but the shape of this issue is clear enough to plan around.

Where AI Is the Wrong Tool in a Firm

While I believe AI is extremely powerful and useful, there are places in a firm where AI does not belong. Anything that amounts to professional judgment or advice of record must stay with the licensed professional; the AI can do the prep but never the sign-off. Anything where a wrong answer carries legal, financial, or regulatory consequence needs a validation step and a human, not an unconstrained model left to its own devices.

The First Step

So, I’ll end where I started. I can’t tell you from the outside which of these solutions fits your firm, in what order they should come, or if you should focus on something else. That requires us spending time together, looking at your business and practices to find the places where AI can earn its keep, and to tell you where it would not.

Next time I will take the same question to another industry and ask again where AI fits and where it doesn’t.