This summer, a team at E7 spent June through mid-September building three AI agents on Superhuman Go. We built them for ourselves, using our own contracts, project records, and internal processes.
A demo tells you what a platform can do. Running it against your own records, with your own people, tells you what it takes to make it useful. This was a deliberate investment in our team’s capabilities, our own operations, and ultimately the services we deliver to clients.
Choosing Where to Invest
We partner with Superhuman, and we’re always looking for ways to strengthen our capabilities and the solutions we bring to clients. This project was an investment in both, with the added benefit of improving how E7 operates.
We sent a questionnaire to every department, then met with each team to understand the problems behind their answers. We were looking for challenges that crossed multiple functions, opportunities to improve how we work, and problems we believed our clients would recognize.
Five candidates emerged. We selected three, with a deliberate emphasis on growth. At least two of the agents needed to support revenue opportunities, alongside improvements to internal efficiency.
The result was three working agents, each addressing a different part of how we run our business.
The Three Agents We Built
SOW Agent
A statement of work defines an engagement: the deliverables, the approach, the commitments, and what’s included or excluded. Consultants, engineers, project managers, and solution engineers all need that information, often for different reasons.
Our SOW Agent helps our teams understand those agreements without searching through lengthy documents or tracking down the person who wrote them.
It provides answers grounded in the SOW, with supporting contract language and links to the source document. It also helps clarify what’s in or out of scope, helping ensure nothing gets overlooked during the engagement.
The timing matters. A question that previously might have taken a day or two to research can now be addressed while preparing for a client conversation, or even while that conversation is happening.
A person still owns the decision. The agent makes the evidence easier to find and evaluate.
Similar Engagement Agent
When we estimate new work, we often look for comparable projects. The challenge is finding the right examples and understanding what actually happened, particularly when the people who delivered those projects aren’t involved in the new opportunity.
Our Similar Engagement Agent searches past engagements and identifies similarities and differences. It compares estimated effort with actual effort, measured in hours rather than dollars.
We deliberately wanted the differences called out. Two projects can look similar on the surface while having very different technical or operational requirements.
For a solution engineer who wasn’t involved in the earlier work, the agent provides access to the experience behind an estimate. That knowledge no longer depends entirely on finding someone who remembers the project.
Better-informed estimates help us pursue the right opportunities, make realistic commitments, and approach new work with a clearer understanding of what delivery will require.
Client Opportunities Agent
Some of the most valuable opportunities to help a client emerge after an engagement is underway.
They might appear as a recurring problem, an unresolved blocker, or a requirement that wasn’t part of the original scope. They’re not always obvious, particularly when several teams are involved in delivery.
Our Client Opportunities Agent reviews active engagements each week and posts a summary into the project’s internal channel. It highlights potential opportunities to help the client, along with project progress, risks, and blockers.
The goal is to help our account and delivery teams recognize needs earlier, while there’s still time to have a useful conversation with the client.
This agent raised some different challenges. An agent that responds to a question gives someone the opportunity to evaluate its answer before acting. An agent that publishes information on its own requires more care around accuracy, permissions, and accountability.
That distinction shaped how we approached the build.
What We Learned Along the Way
One of our most useful discoveries had little to do with AI.
While building the SOW Agent, we brought together years of statements of work that had never been examined in one place. We found terms that meant one thing to one team and something slightly different to another.
Those inconsistencies weren’t obvious when each document was read on its own. They became apparent when an agent needed to interpret them consistently across engagements.
We’re using what we found to standardize definitions and make future statements of work clearer. That’s an improvement to our business regardless of how much time the agent ultimately saves.
The project also reinforced three lessons.
Someone has to stand behind the answer.
One of our engineers made this point during the build. You can’t rely on “the AI told me” when a decision affects a client, a contract, or a project.
That’s why the SOW Agent provides its sources and reasoning. It can also say when there isn’t enough evidence to answer. The person making the decision needs to be able to check the answer and understand its limitations.
The quality and relevance of your records matter.
An agent works with the information you give it, including the inconsistencies and gaps already present.
We restricted our agents to relevant, tagged documents rather than giving them unrestricted access to entire repositories. The Similar Engagement Agent focuses on recent projects because our tools and delivery methods have changed enough that older engagements aren’t always useful comparisons.
Those decisions required understanding the work, not just configuring the technology.
Permissions need to be addressed from the beginning.
An agent should only return information the person requesting it is authorized to see.
That becomes particularly important when agents search across systems or publish information into shared channels. Access controls, source permissions, and the intended audience all need to be considered during the design.
These questions are easier to address before an agent is running than after people have begun relying on it.
Where This Experience Takes Us
The three agents gave our team deeper practical experience with the decisions involved in building AI into everyday work.
Which questions are worth automating? Where does the information live? How reliable is it? Who should have access? And how can someone verify an answer without repeating all the research themselves?
That experience strengthens our own capabilities and the work we do for clients.
The same approach applies well beyond our documents.
Our SOW Agent works with statements of work. Another organization might need answers from project charters, service agreements, feature requests, or user stories.
The documents change, but the underlying need is familiar. People want consistent answers, supported by information they can trust, without searching through multiple systems or asking several colleagues.
Much of that information already exists. Making it accessible, reliable, and useful is where the work begins.
Looking Beyond Time Savings
Time savings are often the first reason organizations consider AI agents. They’re relatively easy to understand and measure.
But we’re interested in the opportunities they can help uncover as well.
Consider what happens when we scope a project. We document what’s included and what isn’t. The excluded work becomes part of the record, and everyone moves on.
Months later, the client’s circumstances may have changed. Something that was out of scope could now be worth revisiting. The information is already there, but nobody is routinely looking through old agreements for those opportunities.
We’re testing whether an agent can review those records and identify items worth bringing back to an account team’s attention.
The potential value extends beyond answering a question faster. It could mean recognizing an opportunity to help a client that otherwise goes unnoticed.
We haven’t established what that will produce yet. We’re testing it the same way we approached the first three agents: using our own information, evaluating the results, and learning from what we find.
Where We’d Start
If you’re considering AI agents for your organization, start with a question your people answer repeatedly.
Where does the answer live? Who needs it? Who is allowed to see the information? And what would change if a reliable, well-supported answer were available in seconds instead of hours or days?
Those are the questions we started with at E7. They helped us identify where agents could be useful and exposed improvements we hadn’t anticipated.
If you have a similar challenge, we’d be happy to work through it with you. Contact our team to start the conversation.
