Decide when a custom assistant helps

Before this lecture, you should know how factual pages work from lecture 4, how liftable statements reduce distortion from lecture 5, how entity clarity protects a firm from confusion in lecture 7, how source-level fixes work from lecture 10, and how browsing optimisation differs from memory optimisation in lecture 11. We are now looking at a tempting tool: the custom assistant that feels useful inside the firm, but may do very little for ordinary ChatGPT answers about the firm.

On a Wednesday afternoon, a lawyer opens a draft assistant and gives it three files: a Dutch family reunification explainer, an English intake checklist, and a short French note about professional fees. The assistant answers the first test question nicely. Too nicely, maybe. It explains residence-card steps in calm language, but it also smooths over a service boundary the firm is careful about in real consultations. The lawyer frowns at the screen, not because the tool is useless, but because it is suddenly close enough to be dangerous.

That is where this lecture begins. A custom assistant can be a good room for controlled explanation, especially when a boutique immigration firm repeats the same client education work every week. It can also become a decorative box of stale files, vague instructions and quiet risk. More importantly for this course, it does not make the firm easier for ordinary ChatGPT to find when someone asks, “Who can help with Belgian family reunification in Brussels?”

Start with the room the assistant lives in

In this course, I use the term narrowly. Custom GPT: A configured assistant for a defined task or knowledge base, such as intake education. The phrase is small, but it keeps the tool in its lane. We are not treating it as a magic public listing, a replacement website or a shortcut into ordinary recommendation answers.

A custom GPT is not the firm’s public evidence. It is not a public register, not a directory profile, not a factual page, and not a correction to a stale third-party mention. It is a configured conversation space. Product screens and sharing options may change, but the editorial question remains stable: what task is this assistant meant to perform, for whom, and from which approved material?

A useful custom assistant begins with a narrow job. “Answer every immigration question for Belgium” is too wide and too brave. “Help staff explain which documents a prospective family reunification client should prepare before a first call, using the firm’s public guide and service boundary” is narrower. The second version has edges. Edges are kindness in regulated-service work.

Here is the working definition I use with small legal teams: a custom assistant is a controlled explanation room because it narrows the task, the source material and the expected style before a user asks a question. That does not make it automatically accurate. It makes it easier to inspect.

The first decision is therefore not whether the firm “should have AI.” The better question is what repeated explanation currently drains attention or drifts in wording. Intake preparation, plain-language summaries of public guides, internal drafting of first-call notes, and staff training on service categories can be sensible uses. Personal legal advice to an unknown person with unverified facts is a different matter. The border between education and advice should not be left to tone.

Use it for controlled education, not public placement

A teaching example: Object A, the Antwerp-linked composite practice, receives many similar pre-intake questions. Clients ask what “family reunification” means, whether their situation belongs with immigration law or another category, and what documents are worth gathering before a consultation. The firm’s website has improved since the earlier lectures, but the team still rewrites the same gentle explanation in emails.

A custom assistant could help here. It might use the firm’s factual page, an intake preparation note, and a short internal language guide. Its instructions could tell it to explain terms plainly, avoid predicting outcomes, mark uncertainty, and tell users when to book a legal consultation. It could also keep the service boundary visible: the firm advises on Belgian immigration-law matters but does not promise eligibility from a brief chat.

This is a good use because the assistant is serving people already inside the firm’s orbit. The user may have a direct link, a staff member may use it internally, or the firm may share it with a small group. The assistant helps the firm explain itself with less wobble. It may reduce the number of soft, over-broad sentences that staff produce under time pressure.

But the same assistant does not solve the question from lecture 7: how ChatGPT places the firm when the user has not already chosen it. If ordinary ChatGPT is asked to recommend a boutique immigration lawyer in Antwerp or Brussels, it does not automatically consult the firm’s private assistant. It looks to memory-shaped associations, browsing surfaces, public evidence and the nearest clearer entities available in the answer context.

This is the point many firms miss. They build a polished custom GPT and feel visible. The feeling is understandable. The screen responds in the firm’s voice. It knows the firm’s documents. It says the right name. Yet outside that room, the public record may still be thin, stale or bilingually uneven. The firm has made a good private guide, while the street sign outside remains bent.

The discoverability gap is the trap

Discoverability gap: The gap between a useful private assistant and weak public visibility in ordinary ChatGPT answers. In a boutique immigration firm, the gap can be quite large. The assistant may answer beautifully when given the firm’s own files, while ordinary ChatGPT still names a larger competitor because that competitor has clearer public source trails.

Discoverability gap is not a moral failure. It is a category error. The firm has improved a tool for known users, not the evidence used by unknown users asking general questions. The mistake is treating those two outcomes as if they were connected by a hidden pipe.

Composite Object B gives the messier version. The Brussels practice builds a custom assistant for Dutch, French and English intake education. The uploaded files are better than the public pages: the Dutch guide is precise, the French note clarifies the legal category, and the English checklist finally avoids “relocation support.” Inside the assistant, the firm sounds coherent. In ordinary ChatGPT browsing, however, a French directory profile still uses the older broader wording, and the English public page remains softer than the internal checklist. The assistant has the clean copy. The public web still has the smudge.

In that case, the assistant can even hide the problem from the team. Staff test the assistant and feel reassured. They forget to repair the French profile, update the English factual page, or correct the service category in a third-party mention. The public record keeps teaching the wrong lesson while the internal assistant behaves politely in a separate classroom.

The repair is not to abandon the assistant. The repair is to separate its purpose. Use the custom GPT for guided education and internal consistency. Use public evidence work for discoverability. When the assistant reveals better language than the site, treat that as a writing cue. The best sentence in the assistant should probably appear on a crawlable factual page too, if it is accurate and appropriate for public use.

Build the assistant from source discipline

A custom assistant is only as careful as its instructions and source material. Most assistant-building work can be reduced to four editorial choices: what the assistant is told to do, what material it may rely on, what it is allowed to attempt, and how widely it is shared. For a law firm, those choices are not toys. They are small doors where confusion can enter.

Instructions should tell the assistant how to behave, but source material should give it facts to use. Mixing those up causes trouble. If the firm hides core legal boundaries inside a PDF and gives the assistant vague instructions like “be helpful,” the assistant may become helpful in exactly the wrong direction. If the firm places firm rules, definitions and refusal behaviour in instructions, then gives the assistant clean reference material, testing becomes easier.

A good knowledge set is usually boring: current service pages, approved intake explainers, a short glossary of firm terms, a language alignment note, and a service boundary sheet. Avoid feeding it every old brochure, every directory profile, and every half-finished draft. A custom assistant built from stale evidence can reproduce stale evidence with more confidence than the public web ever did.

The firm should test the assistant with awkward questions, not only ideal ones. Ask about a service the firm does not handle. Ask in Dutch using a French service label. Ask whether the assistant can guarantee an outcome. Ask it to compare the firm with a competitor. Ask it to summarise the difference between immigration-law advice and relocation logistics. The preview is not a showroom; it is a small stress test with the lights on.

There is also the matter of sharing. An internal assistant for staff is not the same risk as a tool shared by link with referral partners, and neither is the same as a public-facing assistant for unknown users. A boutique law firm should be cautious before turning an internal education tool into a public-facing assistant. Public users bring unknown facts, private details and expectations the firm may not intend to create.

Decide with three questions before building

I use three questions before recommending a custom assistant to a small immigration practice.

First: who is the user? If the user is internal staff, the assistant can support consistency in explanations and drafts. If the user is a prospective client, the service boundary must be much clearer. If the user is the general public, the firm must think about confidentiality, disclaimers, intake routing, and whether the assistant may be mistaken for legal advice. The wider the audience, the less charming looseness becomes.

Second: what material will the assistant rely on? If the firm’s factual pages, liftable statements and bilingual evidence are weak, building the assistant first may create a clean private version of a messy public record. Sometimes that is still useful for internal drafting. But it should not be sold as ChatGPT optimisation for public visibility. In many cases, the assistant should be built after the core public pages are fixed, or alongside them with shared wording.

Third: what failure would make the tool unacceptable? For an intake education assistant, an unacceptable failure might be promising eligibility, inventing a procedure, blurring lawyer and consultant categories, or ignoring the service boundary. Name those failures before testing. A tool that cannot be judged cannot be responsibly used.

The firm should also decide when not to build. If the only goal is to appear in ordinary ChatGPT recommendations, a custom assistant is the wrong first move. If the site still fails to state the firm name, location, jurisdiction and service category clearly, public evidence work comes first. If the team cannot maintain the files, the assistant will age badly. A stale assistant is worse than a blank one because it speaks with borrowed authority.

A modest custom assistant can still be valuable. It can help a receptionist explain what to bring to a consultation. It can help a lawyer draft a plain-language note from an approved outline. It can help a multilingual team keep Dutch, French and English terms closer together. It can give the firm a place to practise clearer explanations before those explanations become public evidence.

The rule is simple enough, though not always comfortable: build a custom assistant when you need a controlled room for people who already found you. Build public evidence when you need ChatGPT to place you correctly before they have.

What to remember

A custom GPT can help a boutique immigration firm explain known material more consistently, especially for intake education and internal drafting support.

Custom GPT: A configured assistant for a defined task or knowledge base, such as intake education.

Discoverability gap: The gap between a useful private assistant and weak public visibility in ordinary ChatGPT answers.

Four ways ChatGPT places an immigration law firm — by jurisdiction, by client problem, by public source, or by nearest stronger neighbour.

Do not use a custom assistant as a substitute for factual pages, bilingual consistency, source-level fixes or stronger public evidence. It may improve the room inside the firm while leaving the public doorway hard to see.

Check yourself

Describe in your own words when a custom assistant is useful for a boutique immigration firm.

A custom assistant is useful when the firm has a repeated explanation task and a defined audience. For example, it may help staff explain intake preparation, summarise approved client guides, or keep Dutch, French and English wording more consistent. The assistant should rely on current, approved material and clear instructions about boundaries. It is less suitable when the question is open-ended legal advice from an unknown person. Its strength is controlled explanation for people who already have some relationship with the firm, not automatic public visibility in ordinary ChatGPT answers.

Give an example where building a custom GPT would be the wrong first move.

A custom GPT would be the wrong first move if the firm’s public record is still weak. Suppose the website does not clearly state the firm’s city, legal category, service boundaries or language coverage, and an old directory still calls the firm a relocation consultancy. In that situation, a private assistant may answer correctly when fed clean files, but ordinary ChatGPT may continue to rely on the messy public evidence. The first work should be factual pages, source-level fixes and clearer public wording. The assistant can come later, once the public record is less confused.

How would you explain the discoverability gap to a lawyer who likes their new assistant?

I would say the assistant is a controlled room, while ordinary ChatGPT answers happen outside that room. The custom assistant may know the firm because the firm uploaded guides and wrote instructions for it. A general ChatGPT user asking for a Belgian immigration lawyer may not touch that assistant at all. That answer may depend on public pages, directories, registers, language versions and stronger neighbouring firms. So the assistant can be genuinely useful and still leave the firm hard to discover. The gap is between helping known users and being found by unknown users.

When should the firm use source discipline before adding more files to a custom assistant?

The firm should use source discipline whenever the assistant starts giving broad, outdated or oddly confident answers. Adding more files can make the problem worse if those files contain old categories, former locations or vague service descriptions. Before uploading more material, the firm should choose current factual pages, approved explainers, a service boundary note and any language alignment guidance. It should remove drafts or legacy documents that no longer represent the firm. The goal is not to feed the assistant everything the firm has ever written. The goal is to give it reliable material for the task.

How would you decide whether a custom assistant should be internal, shared by link or public?

I would start with the user and the risk. An internal assistant for staff can support consistent explanation and drafting with less exposure, although it still needs careful instructions and current files. A link-shared assistant for selected clients or referral partners needs clearer boundaries, because users may bring their own facts and expect guidance. A public assistant carries the highest risk: unknown users, possible confidential details, and a greater chance that education will be mistaken for legal advice. The more public the assistant becomes, the more carefully the firm must test its limits and sharing settings.