Professional services solutions provider
Thad Jampol, co-founder and chief product officer of Intapp, said the AI specializes in the kind of judgment-intensive work particular to professional services providers like CPA firms that traditionally have been difficult to automate at scale. This development, he said, is "the biggest and most important thing we have ever built."
"There are requirements specific to your industry that generic AI and generic agents just don't have. It's not built in. They don't address it. Your clients, your ethical walls, your deal history, your compliance logic, your independence rules that institutional knowledge, those relationships aren't built into any general-purpose AI. And that's the problem we set out to solve," he said during his presentation.
Celeste is a standalone product for professional services firms that performs practice management tasks such as deal screening, conflict clearance, engagement management, business onboarding, cross-selling, work-to-cash management and more. Overall, they encompass multi-role, multi-departmental workflows that go beyond simple rules-based processes.
Playbooks, skills, connectors
The platform centers around a concept called Playbooks, which represent an individual firm's unique methodology for a given task. They can be thought of as blueprints that codify a firm's best practices and workflows. Each Playbook is made up of several skills, which are specialized capacities the AI can execute, such as research analysis, financial modeling or competitive intelligence. Celeste uses these skills to execute a Playbook, drawing on connectors to access the necessary firm data (plus, if the user so chooses, third-party information from credit rating agencies) to complete the task.
"Skills are what Celeste can do; and Playbooks are what your firm should do, the actual workflows," said Roger Smith, CEO of Term Sheet (which was bought by Intapp last year) during the conference.
He cited the example of Celeste being used to guide dealmaking. Every week, he said, a firm receives dozens of potential deals, mainly from investment banks and other intermediaries. Screening typically is done by a team of humans who read confidential memos, pull comparisons and draft summaries, a process that normally takes days due to the amount of judgment required. Much of this judgment, which lets a senior partner know whether or not something is a potentially attractive deal and should be pursued further, took years to develop over the course of a long career. Celeste, he said, can replicate and automate these years of experience via Playbooks.
During the presentation, he showed that Celeste was running the deal screening Playbook overnight and processed this week's incoming opportunities. For each one, Celeste assessed the fit against firm-specific criteria, sector focus, deal size range and return thresholds, which were all pulled from Intapp Deal Cloud.
"But here's where it gets even more powerful. It also cross-referenced against our complete deal history, not just what we invested in, but why we passed on similar opportunities. It knows that we passed on a health care opportunity in 2023 because of customer concentration risk. It surfaces the pattern recognition, normally locked in senior partners' heads and trapped across scattered memos. We're not just saving analysts time. We're doing something that they never could. We're activating decades of institutional knowledge at scale," he said.
Celeste, he said, flagged two deals as a strong fit and worth a second look, and the rest were determined to be passes (with an explanation why). From there, he brings the recommendations to his supervisors to validate the information and assess the plan. Smith added that Celeste does not just provide insights but helps plan as well.
"Celeste isn't just waiting for me to ask what to do next. It's already prompting me. Do I want to pressure test this deal against our return targets with a preliminary LBO model? That's exactly what I need to do before I take this to my partner. Let's go ahead and do it. Here's the complete model built out from our own template, properly structured, multiple analytical tabs and instructions to guide us, ready to work. It used to take days, screening the CIMs, validating the fit, running the numbers Celeste compressed into minutes. And everything Celeste produced is traceable," he said.
Users can also make their own Playbooks from scratch. For instance, someone could built a Playbook instructing the AI to proactively monitor the market for M&A activity and, when found, flag sponsors, exit windows and add-on opportunities for their portfolio, and to do all this continuously in the background. He said users describe what they need the Playbook to do much as they would brief a new staff member.
Jampol, in a later interview, said Celeste comes with over 1000 pre-built Playbooks, including "about two dozen" specific to the accounting profession that include things like independence checks, new client onboarding, and engagement management. He added that there's also "several dozen" skills and "that list keeps growing," but right now they include capacities like extracting EBITA from a P&L client, exposure risk scoring, revenue and realization analysis and regulatory classifications. Understanding the stakes of the tasks these skills would be applied to, he noted that many can be run deterministically versus generatively.
Creating new Playbooks does not require any special coding knowledge. He said they are created 100% in plain English. People instruct the AI as they would a colleague, and the AI is meant to then figure things out from there.
"You tell it what you want as you would a colleague: 'Here's what I think you should do, here's what we want to accomplish, and you figure out how to do it. Maybe I have some ideas.' And it will come back and it will say, 'OK, here's what we're going to do. Here's how we're going to do it. Here's the skills we're gonna use, and here are the connectors.' And then you have an opportunity as the agent creator, to say 'I like that' or maybe I want to tweak this over here. But then you publish it, and off it goes," he said.
To do all this, Celeste draws on the entirety of the user's enterprise data pool. This includes (but certainly is not limited to) emails, documents, pitch decks, cloud drives, partner notes and meeting transcripts. It also includes third party data sources like Moody's or Bloomberg (if they chose to connect them), and government information like sanctions lists or regulations. He said it can even monitor "adverse media" and "time card narratives" 24/7 to detect and act on material changes to the engagement that introduce new risk to the firm.
"It's how Celeste knows that a qualified opportunity at your firm is different from a qualified opportunity at the firm down the street," he said.
This does not just apply to business opportunities. When connected with Intapp Time, for example, the company's timesheets solution, it can also analyze staff activity. Laura Saklad, the company's lead for the legal industry, asked the audience to consider a large law firm whose leaders want to better understand where they are capturing the value of the time worked. By running the Revenue Realization playbook, Celeste was able to show, among other things, the amount of time their professionals actually worked, the hours they recorded on their timesheets, and the hours that made it onto the client invoice.
"And now you can see here, Celeste has highlighted an important insight from this data: my team self-edited over 3,000 hours of work off of their time entries before they submitted them. This surfaces what was really previously hidden, revenue leakage, information that I can now address. Celeste will also put information about what is driving write offs right at my finger tips," she said.
With information like this, she said, firms can optimize staffing and protect margins based not on gut feel but hard data. She added that these insights can also be used to capture that which many organizations have been trying to find lately: ROI on AI.
"How quickly is my firm actually incorporating new productivity-enhancing AI tools into my practice? Let's open the AI impact playbook to see what we can learn, because Time will automatically capture which AI tools are being used by whom and on what work, and then feed that directly into Celeste. I can see here that 15% of the work captured by time was executed using one of my firm's AI tools, and then I can see precisely which tools are being used. I can see this information by role. I can see it by task, and as we saw at the firm-wide level, we can drill into this information to understand which tools precisely have the greatest adoption and which do not … Now you can measure the real ROI of your AI investments," she said.
Governance and safety
Beyond absorbing the firm's individual methodologies, Celeste takes into account the firm's individual governance structure as well.
"Everything Celeste does respects your firm's heightened professional compliance and confidentiality obligations. Ethical walls, insider lists, independence rules and market abuse regulations are enforced automatically at the architecture level. It is not bolted on. This is built in as a fundamental design principle," said Jampol.
Elaborating further during the later interview, he said Walls can be thought of as a policy management engine. Firms load a list of all their different policies impacting clients, engagements and staff, which include information barriers, ethics requirements, independence and more. He added that in cases where there are multiple policies overlapping (which happens more than one might think), the AI reads how the firm itself prioritizes things and moves from there. These policies are kept up to date across all engagement teams.
Another major component is in Celeste's API. Jampol described it as a "mother-may-I API" where, if someone wants to access a piece of client or engagement information, "it will say, yes, you can access that or no, you cannot."
There are other companies that were very confident in their respective AI controls who went on to have embarrassing public incidents with their models. Jampol was aware of these examples and said what makes Walls different is that controls are not a matter of asking the AI politely to please not violate them — policies are enforced at the architecture level, so even if the AI wanted very badly to circumvent the controls placed upon it, there would be no actual way to do so.
"These guardrails are not self-imposed by our AI. We're not asking the AI nicely to not go do something. We're actually locking the door from outside and saying, 'There is no way you're going to be able to get there even if you tried.' All of this security, governance, confidentiality is enforced at the infrastructure layer outside of the model. … And so the controls that govern what Celeste can access, or what actions it can take, or which walls separate one client information from another are enforced at the data and the permissions layer. It doesn't matter what Celeste decides. There's no way it can circumvent any walls. It has no architectural access," he said.
When it comes to controlling the tendency for AI to make up information (i.e., hallucination), he said there are three core approaches. The first is through restricting the context window to just the information that is relevant, as irrelevant data increases the risk. For example, when discussing a specific engagement, it can restrict context to only that which is germane to that specific engagement. Then, if that fails, there are what he called "knowledge agents." These are agents that manage engagement files, which give them a vast store of information specific to the firm that serves to take the inferential load off the execution models. And finally, the AI uses deterministic calculation for tasks where accuracy is absolutely critical, versus the more probabilistic approach of a typical AI. Still, when asked whether he was confident there was zero possibility of hallucination, he said no one can guarantee that.
"We don't have the hubris to say we're going to have zero hallucinations. I think that's not a thing that anybody should be saying right now. That being said, we take it very seriously," he said.
Product revamp
While Celeste is billed as a standalone product, Intapp plans to revamp its entire application suite to revolve around it. Already it is embedded directly into Deal Cloud, Conflicts, Intake and Time. More applications will follow, with the compliance and collaboration applications slated to be next, but overall the company expects all revisions to be done by the end of spring, according to Jampol.
Those who already own the premium package for DealCloud, Conflicts, Intake and Time will get the Celeste capabilities added to those products at no additional cost. However, there will be consumption-based pricing considerations for usage. Jampol said the pricing of the premium packages will not be impacted by the addition of this AI.
He added that Intapp intends to continually improve Celeste, especially where it concerns Playbooks. The company is currently looking for design partners with good use cases that could be worked into Playbooks. To this end, he said, Intapp is also deploying applied AI engineers to certain customers to help develop them. This is in service of the overall goal of assisting organizations like CPA firms with the ability to run the actual business of the firm more smoothly.
"We're really focused on the business side of the firm," he said. "We're not doing the audits, we're not doing the tax returns, we're not generating the legal briefs and doing the contract reviews. What we're focused on is: How does the firm grow? How does it win new clients? How does it grow those clients? How does it bring other services into that client, the cross-serving, cross-selling? How do we run very sophisticated cross-border, global network, independence, optimized, business acceptance? How do we help you optimize your engagement margins? How do we help you analyze and optimize the balance of trade across a global network? This is what we mean when we say the business of the firm. It's these types of things."
