The firms rolling out AI vaults this year are largely building them in Harvey or comparable platforms, scoped for the practice of law. Those deployments serve the attorneys. Nothing equivalent has been built for marketing and business development, and nothing appears to be coming.
That leaves marketing teams in an awkward position. The underlying advice is sound. The vault, the second brain, the context layer, whatever a firm settles on calling it, is what makes AI produce work that sounds like the firm rather than work that sounds like nobody. Guy Alvarez recently made the case that firms should build that layer before buying another tool, and the sequence he describes is correct. Across thirteen sessions at ILTACON this year, nearly every one addressed some version of it.
The difficulty appears when the instruction reaches the person responsible for carrying it out.
They open a folder and begin listing what would need to go in it. The practice descriptions the partners argued over and signed off on. The RFP responses that won, and the ones that did not. The style rules nobody has written down. The matters cleared for public use and the ones that are not. Somewhere around the third category, the scope of the thing becomes clear, and the folder closes.
Nothing gets built. Not because the team isn’t capable, but because the team itself is already overwhelmed. Executing on a project at this scale takes the entire team’s effort, and they are already overtasked.
Most writing on context layers assumes a team with the capacity to document its entire operation. A team of two does not have that capacity. Neither does a team of thirty doing the work of forty-five, which is closer to what By Aries encounters at most firms. The marketing director is already covering strategy, execution, the website, and a steady queue of attorney requests, and has recently been handed AI readiness because firm leadership read something and decided it was a priority.
Handing that person a documentation project with no defined endpoint is a reliable way to ensure it never begins.
So marketing teams need to build their own, but many think it must be comprehensive, when it can start small and be hyper-focused on something truly valuable.
What a context layer is
A prompt library is a collection of prompts that your team built, tested, and found to be effective at achieving their intended goal.
A context layer is where that information lives permanently. Once it is in place, the prompt becomes short. “Reference our last three healthcare RFP responses and draft the cover letter” is the entire instruction. The context does the work the prompt used to do poorly.
The concept is straightforward. The difficulty is deciding what belongs inside it, which is where most teams stall.
Begin with one workflow
Rather than trying to capture a department’s collective knowledge, teams should choose a single workflow that carries real weight and assemble only what that workflow requires.
By Aries asks teams four questions when they are choosing:
- Is it run monthly or more often?
- Would AI actually help this workflow? Meaning, would using AI actually improve this workflow?
- Is it obvious what done looks like? Meaning there is a finished, approved version the team could point to.
- Does the team produce different quality each time, depending on who picks it up?
Four yeses identify a candidate. The third question carries the most diagnostic weight, because inconsistent quality is the clearest signal that the knowledge governing that work lives in someone’s head rather than on paper.
Here are the most common workflows:
- Email management and triage
- RFP responses
- GEO-optimized attorney bios
- Lateral partner announcements and the communications chains that follow them
- Distribution and promotion of thought leadership
- Awards and submissions management
- Reporting and communications regarding the team’s capacity, availability, and accomplishments
These are just a few of the workflows By Aries has seen teams build.
Each is Iterative, has a defined finished state, and is executed differently depending on who has capacity that week.
The questions matter as much for what they rule out. A firm anniversary campaign fails all three: it runs once, there is no prior approved version to point to, and consistency is not the problem. A website redesign fails for the same reasons. So does a one-time rebrand, a merger announcement, or an office opening.
Those projects are usually the most visible work on the team’s plate, which is exactly why teams reach for them first. They are also the worst possible place to start. A context layer earns its cost through repetition. Building one for work that happens once produces a folder nobody opens again.
What a single workflow requires
Let’s look at four of the workflows mentioned above and break them down into the five or six documents needed to support them, most of which already exist somewhere in the team’s files.
| Workflow | What the context layer needs | Where it usually already lives |
|---|---|---|
| RFP responses | Approved practice group descriptions, the last three responses that earned a reply, the response template, rate and pricing structure, cleared representative matters | Proposal folder, prior submissions, the BD manager’s drive |
| Attorney bios for GEO | Current approved bios, practice descriptions, the firm’s bio format and length rules, the language the firm uses for credentials and outcomes, cleared matters by attorney, the standards the firm is using to assess GEO | Website CMS, the bio template, directory submissions |
| Lateral announcements | The announcement template, prior announcements the firm was happy with, practice group positioning, the approval chain and who signs off, the standard set of channels | Comms archive, press release folder, the marketing director’s memory, the team’s project management tool |
| Thought leadership distribution | The alert or article itself, channel-specific formatting rules, attorney voice samples, the firm’s disclaimer language, the target audience for each practice | Content calendar, prior social copy, the style guide only the comms manager references |
Focusing the build of your context layer transforms a monumental task into an achievable one. Most teams could assemble one of these folders in an afternoon once they stop attempting to do it department-wide.
The overlap in the middle column is the point. Approved practice group descriptions appear in three of the four. Cleared representative matters appear in two. The second workflow a team builds reuses much of what the first one gathered, and the third costs less again.
The layer compounds as it grows. It doesn’t start as a layer; it builds one workflow at a time.
The case for marketing leadership
For those leading a marketing function, the argument extends past efficiency.
On most teams today, two or three people have developed their own effective setup. A director with a project they have tuned over months. A business development manager with a prompt they will not modify because it finally produces what they need. A coordinator producing the strongest content on the team because they found something that works.
None of it is documented. None of it is shared. When one of those people leaves, the capability leaves with them.
Instead of creating real AI capability, your team has built individual productivity that didn’t scale to the institutional level. Worse, it doesn’t survive the inevitable personnel changes every firm faces.
One documented workflow the entire team can run is worth more to the firm than three excellent private ones.
Curation is the work
The consideration most often omitted from this conversation is also the one most likely to cause damage.
Knowledge management practitioners at ILTACON 2026 named it directly this year. A large language model treats every document it is given as equally authoritative. A first draft, a redlined version, and the final approved copy all read as equally authoritative to AI. It has no mechanism to determine which version is the final approved version, so it draws from all three drafts.
Most legal marketers have encountered the smaller version of this problem. A lawyer reviews a pitch deck and says that is not what happened on that matter. The language came from the experience database. The database was stale, or the lawyer never updated the record, and the error has now reached a client.
The failure to maintain and update the context layer scales with its size.
A context layer is therefore not an archive. Archives are complete. A context layer is curated, and the curation is the work. Only final, approved, current material belongs in it. Material struck in review stays out. Every document carries a status and a date, because AI will not infer either.
In practice, this requires separating what AI reads from what AI produces. The structure By Aries uses with client teams has three tiers, and the access rule attached to each one is what does the work. A strong orchestration prompt governs it all, often called Project Instructions in Claude, ChatGPT, and Perplexity, and Notebook Instructions in Copilot.
| Tier | What lives there | Access rule |
|---|---|---|
| System | Approved practice descriptions, winning responses, templates, style and voice rules, governance and confidentiality lines | Read-only. Nothing is written here by AI, and nothing enters without review |
| Inputs | The live RFP, interview notes, deal sheets, research, the raw material for the task at hand | Read and write. Contents are temporary and tied to a specific job |
| Outputs | Drafts, pitch decks, bios, alerts, whatever the workflow produces | Write destination. Nothing here is treated as authoritative until it is promoted into System |
The maintenance of the context layer often gets skipped, and skipping this key step is how the problem recreates itself. It is how high-quality work becomes AI slop and a marketer or BD professional’s credibility gets crushed. A draft that gets approved does not belong in Outputs indefinitely. Somebody has to move it into System, stamp it with a date, and retire whatever it replaced.
Without that step, reference material and working drafts blend within a month, and the team is back to an AI that cannot tell approved work from abandoned work.
Documents are only half of it
Two teams can load identical documents and produce entirely different output. The difference between the two teams is not what is in the context layer. Instead, what separates the two is the team’s ability to encode their judgment and standards into AI through skills that apply to the outputs AI creates.
Documents tell AI what a firm has. They do not tell it what good looks like.
The clearest illustration of this error the By Aries team sees involves soft hallucinations. For example, AI produces an output that includes a line like this one: “across the industry, we are seeing,” with no cited source behind it. It is not a hallucination in the conventional sense, but rather a soft hallucination. It is a confident claim with nothing backing it, written in the language a lawyer would typically use. And most legal marketers miss this type of hallucination.
In the AI fluency assessments By Aries runs with legal marketing teams, that is the single most-missed item.
No quantity of documents corrects it. A written standard does. One line is sufficient: every claim about industry conditions, trends, or conversations carries a named source, or it is cut.
That is what a skill is. Not a prompt pasted in each session, but a standard the system applies to every output.
Written out, a first one is short enough to fit on an index card:
Sourcing standard
Every claim about industry conditions, market trends, or what clients are asking carries a named source, a date, or a specific matter it came from. Claims that cannot be sourced are cut, not softened.
Do not write “across the industry,” “many firms are finding,” “clients increasingly expect,” or “we are seeing” unless the sentence names where the observation came from.
When a claim is accurate but the source is confidential, write it from the firm’s own experience instead. “In our work with healthcare clients” is defensible. “Across the industry” is not.
This one paragraph of text created as a skill eliminates the most common failure in AI-assisted legal marketing content.
Skills also let the firm encode judgment in other ways. For example, how the firm name appears on first and subsequent references. What the firm never promises in a pitch. The tone difference between a client alert written for a general counsel and the same update written for a founder. The outcome language that requires a qualifier. The judgment calls a junior person would not know to flag.
A useful test for identifying these: think of the last time someone’s draft came back with the same redlines as last week. That correction is the standard waiting to be written.
Seasoned team members hold these standards internally, which is exactly why quality shifts depending on who handles the work.
Writing it down lets a small team perform like a more experienced one.
What remains human
AI has the capacity to assist and even execute on a lot of the tedious work many marketing and business development teams do daily. But four core areas still require a human in the loop.
- Strategic decisions: What to publish at all, which topics matter, when to skip a cycle entirely.
- Brand voice nuance: The line between sharp and too sharp, and whether a position actually sounds like the firm.
- Relationship calls: Whether to push back on a partner request, how to position a lateral hire, what is communicated directly to a client.
- Crisis communications: When the firm is in the news for the wrong reasons, the voice is human every time.
Where it is unclear whether AI belongs in a step, the default is human.
Start smaller
AI can feel like a race to a finish line that keeps moving. When ROI is impossible to measure and making the case for FTEs feels like a monumental ask, investing in the core components that ensure the AI tools your team is using are actually being used and producing good outputs is the only strategy that feels defensible.
So, encourage your team to start by selecting one workflow, gathering the four or five documents it requires, documenting the standards the firm’s strongest practitioner was already applying, and running the workflow until it is routine.
Then empower them to build the next one.
That is what keeps a team’s use of AI practical rather than aspirational.
