1. Option A: Clean up your core systems of record

Keep your CRM, ERP, and internal databases well maintained and treat them as the single source of truth your agents rely on. This is the right long-term ambition, but in a large enterprise it is rarely realistic to get every system perfectly clean and aligned before you start using agents.

2. Option B: Curate smaller, trusted data sets for agents

Create specific, well-governed data sets that an agent is allowed to pull from for a given use case. You reduce the risk surface and gain more control over quality. You do introduce some double maintenance, but for most large organizations this is the more realistic way to get reliable results from AI agents.

Templafy’s Document Agents help here. They add a layer on top that lets you control which templates, content libraries, and data sources each agent can actually use. That way, the agent can build a deck that is high quality in both content and information, consistent with that specific document workflow, and cheaper to run because you’re not relying on tokens to generate and find everything from scratch.

The technical bit, made simple

Here’s how it works in practice.

The harness

Custom Agents add a management layer, or harness, on top of your preferred AI tools. This harness connects to your structured data (like CRM, ERP, and BI tools) and unstructured knowledge sources (like slide libraries, policy documents, playbooks, and templates), and defines what each source should be used for. It also locks in your brand rules, layouts, and approved content, so the agent only pulls from the right places in the right way.

The orchestration

The orchestration is the playbook inside that harness. It decides which data and knowledge sources the agent can use for each step, which questions it should ask, and when to call which system. It controls how information is combined, how it is structured in each slide, and how it all comes together as a complete, business‑ready document or presentation in one go.

Through Templafy’s self‑service admin interface, your teams set up these harnesses and orchestration flows once per document type, then reuse them across the organization. From there, anyone can ask the agent for “our standard RFP response” or “a QBR for this account”, and the agent knows which systems to pull data from, which libraries to use for content, and how to assemble everything so it already looks and reads like your best internal example, not a generic AI draft. 

Once people see what an agent can do with custom document types, the next question is almost always: “What about all the content sitting in SharePoint?”

At our recent events, I heard it from both extremes: one leader with literal petabytes of unorganized legacy content, another with a beautifully tagged, highly tuned SharePoint setup.

In both cases, the real problem is the same: how do we turn this web of systems and content into something an agent can reliably use for one specific workflow?

The trick is to stop thinking “clean up SharePoint” and start thinking “support this specific document/workflow.” You don’t need to fix every file or system. You need to define the data flow for one use case. I usually break it into three practical questions:

1. Which systems hold the source of truth for this document?

For this one-use case (say, RFPs), the authoritative information is usually spread across CRM, ERP, pricing tools, policy libraries, and specific SharePoint sites or folders.

2. Which fields and content blocks does the document actually need? 

For example: customer name, segment, ARR, product list, contract dates, current risks, latest roadmap slide, standard legal clauses. Name them explicitly.

3. Where should each of those fields land in the document? 

Map them to concrete places: “this field populates the cover page,” “these rows fill the pricing table,” “this policy paragraph goes into section 4.2,” “these KPIs feed slide 3.”

Once you’ve answered those three, “unstructured content” becomes a defined data contract between your systems and your document. You can then point the agent only at those specific data sources and content blocks, and let it assemble the document, instead of people hunting through folders and links every time.

There’s a bonus to doing it this way. When you force yourself to list the actual fields and blocks, you quickly see what is missing, duplicated, or never used. Fixing just that short list often does more for document quality and automation than any attempt to “clean up SharePoint” in the abstract.

This is the big one. When people talk about AI adoption, they usually jump straight to the tech, but the real challenge is change management. You’re dealing with fast‑moving tools, different levels of understanding, and very mixed opinions on AI across the business.

In practice, making agents a success comes down to a two‑part strategy. 

1. First, get specific about what you are trying to achieve. That means answering two questions:
   • Functionally, what do you want the agent to do and deliver at the other end?
   • Operationally, what tangible benefits do you expect to see, for example time saved, higher accuracy, or fewer review cycles?
2. Then, find your internal champions. Don’t try to force a top‑down mandate on a skeptical workforce all at once. Instead, find the teams that share a clear use case and actually want to use the tool. Let them shape and optimize the agent, and let them talk loudly about the results to everyone else.
   
   Internal peer advocacy beats an executive memo every single time. If you want more usage and more buy‑in, your champions need to prove the value first.
   

If London and Copenhagen proved anything, it’s that the appetite for AI automation is huge, but good outcomes are not automatic.

Agents aren’t magic; they are execution engines that only work at scale when they have clean data, clear boundaries, and well‑mapped processes behind them.

Thank you to everyone who attended, challenged us, and shared their insights, and to everyone who made it to the end of this post. I hope it has cut through some of the hype and jargon and made agentic documents feel a bit more real and achievable.