We gave everyone a workspace, but kept one commercial memory

We are preparing MAS.AI to meet a much larger market.

The constraint is familiar: most of the available money should go into product development, not advertising. At the same time, two colleagues are joining the commercial work. Each will use a separate ChatGPT account to research markets, assess firms, identify relevant people and prepare possible approaches.

The obvious solution would be to give everyone the product description and ask them to start finding prospects.

That would scale activity quickly.

It would also create three versions of the market, three versions of the product and three different ideas of what we have learned.

So we chose a different architecture.

Each colleague gets a workspace. Only the shared system gets a memory.

The unit is the firm, not the message

The work starts with an assigned market and a bounded group of firms.

One firm is treated as one commercial unit. One primary recipient is prepared at a time. The colleague's task is not to produce as many names or messages as possible. It is to build a small, inspectable case for why this firm may be worth understanding now.

That case must distinguish between three things:

• what public evidence actually shows;
• what we currently hypothesise; and
• what remains unknown.

Only then do we prepare a possible message.

The message is not the product of the process. It is the final expression of the reasoning that came before it.

Research can be distributed. Authority should not be.

Each colleague can investigate firms, compare observations, identify a plausible primary recipient and draft an approach.

Their ChatGPT workspace can help structure the work and preserve the local trail.

But it cannot choose a market, change the product promise, approve a claim or contact anyone.

Those decisions return to one central commercial master and one human approver.

This distinction matters because separate AI accounts do not share context automatically. Without a deliberate handoff, each workspace slowly develops its own assumptions, vocabulary and memory of what worked.

We therefore end every working round with a fixed transfer package:

• the assignment and its boundaries;
• the firms considered and their recommended order;
• the primary recipient for each firm;
• the evidence and its sources;
• the commercial hypothesis;
• the unknowns and risks;
• the draft message;
• the decisions that still require approval; and
• the learning that may change the next round.

The transfer package is the bridge between distributed work and shared learning.

Benchmark and MAS.AI create different reasons to pay attention

This operating model connects naturally to both L&F CG Benchmark and MAS.AI, but the two should not be collapsed into the same commercial logic.

The Benchmark can make observable differences between firms more useful. It can help a managing partner, business developer or marketer understand how the firm's public commercial capability appears relative to relevant peers.

That creates a legitimate reason for a conversation.

But the Benchmark must not quietly become a sales ranking. Its analytical learning must remain separate from the commercial learning created by outreach. A firm can disagree with an observation without disproving the method. A firm can respond positively without proving that the underlying assessment was correct.

MAS.AI starts somewhere else.

It looks for observable change, signal convergence and possible legal demand. That may create a hypothesis about timing, relevance or fit.

But a signal is not buying intent. A hypothesis is not permission to contact. And a reply does not retroactively turn the original interpretation into fact.

The system therefore needs two learning loops:

Benchmark learning: Did the evidence, definitions and comparisons hold?

Commercial learning: Did the recipient recognise the problem, respond, redirect us or reveal a better route?

Keeping those loops separate protects both products.

The aim is not automation. It is controlled leverage.

With limited resources, broad paid marketing is unlikely to be our advantage.

Our advantage has to come from arriving with more relevance than our size would normally allow.

That means using public evidence to retire generic questions before contact. It means choosing fewer firms for better reasons. It means approaching the person most likely to understand the problem, rather than the person whose email address was easiest to find. And it means making every response improve the next decision.

AI can make much of that work faster.

It can search, organise, compare, challenge, draft and maintain structure. Two colleagues can cover far more ground with these workspaces than they could through unassisted research.

But the leverage disappears if speed produces fragmented product claims, duplicated contact or invented certainty.

The important design choice was therefore not which prompt everyone should use.

It was deciding what must remain common when the work becomes distributed.

What changed

We stopped thinking of the colleagues' ChatGPT accounts as two additional outreach engines.

They became controlled contributor workspaces feeding one commercial system.

What I think matters

Small companies do not need to imitate the campaign machinery of larger competitors.

They can compete by building a better connection between evidence, timing, recipient choice, human judgment and learning.

The scalable part is not the number of messages.

It is the quality of the commercial memory behind the next one.

What remains uncertain

We still need to test whether the transfer packages are concise enough to use consistently, whether different contributors reach comparable conclusions from the same evidence and how much central review is required before quality becomes stable.

We also need to learn how narrow the first market assignments should be.

So the first experiment will be deliberately small: a few firms per colleague, no automatic outreach and human review before any external contact.

We are not trying to automate our way into the market.

We are trying to learn our way into it without losing control of what we know.