Yesterday, I posted about a tension we're working through at InnovAItion Partners. When professional services firms use AI for thought leadership, do their teams want to delegate (fill out a form, get a draft) or collaborate (have a back-and-forth conversation with the AI before it writes)?
Many people said "both." Which is correct, but it also raises a follow-up question that almost nobody is addressing yet: these two modes of working require fundamentally different AI architectures under the hood. And that has massive implications for how firms should be thinking about their AI investments.
Let me walk through what I mean.
Marketing professionals at law firms who need to turn attorney insights into client alerts typically know what they need. Topics are defined, quotes or interview notes are ready, and target audiences and publication channels are clear. What AI provides in this scenario is execution speed.
Traditional AI workflows shine here. Structured workflows take a fixed set of inputs — think of it as a well-designed form — run them through a predetermined sequence of prompts, and produce a polished draft. Every submission of the same request type triggers the same steps in the same order.
Behind the scenes, you're building what's called a prompt chain: step one takes raw inputs and organizes them, step two drafts content in the firm's voice, step three edits for compliance or style guidelines, and so on. Each step is predictable and repeatable.
For agencies, this model is gold. Producing 30 client alerts a month across eight practice groups demands consistency. Your team has already done the creative thinking during the content capture call with the attorney, and the AI's job is to take that thinking and turn it into a well-structured piece as fast as possible.
Delegation works beautifully when the inputs are well-defined upfront, the process stays consistent across every run, speed and reliability take priority, and the person using the tool is working with someone else's expertise.
Now consider a different scenario. A partner at a mid-size firm wants to write a LinkedIn article about trends in private equity healthcare deals. Fifteen years of doing these deals has given her opinions, war stories, and insights that no form could ever capture, because she doesn't even know what's relevant until someone asks her the right question.
Agentic AI enters here, and the architecture looks completely different from what I just described.
An AI agent can loop. After asking a question and evaluating your answer, it decides what to ask next based on what you said, then keeps going until it has enough context to produce something genuinely good. Agents also have access to tools, from searching for recent news related to your topic to pulling in data or cross-referencing your points against your firm's previous publications.
No two agent conversations follow the same path. Mention a deal structure in passing that has interesting implications, and a well-designed agent will catch that and dig deeper. Give a vague answer, and it will push back and ask for specifics.
My co-founder Sally Slater who spent years on the agency side building content for law firms, made a sharp observation about all of this: the agent is essentially doing what a skilled ghostwriter does. A great ghostwriter doesn't hand you a questionnaire. Great ghostwriters interview you. Sally pointed out that the best ones notice when you light up about a topic and follow that energy. Good agents can do the same thing, asking the question you didn't know needed asking.
Collaboration shines when the person in the chair is the actual SME, when each piece of content is unique in structure and angle, and when quality and depth take priority over speed.
Here's where I think a lot of firms are going to run into trouble. Most AI tools built for content creation right now offer one mode or the other. You get a form-based tool built for delegation, or you get a chatbot-style interface that tries to be collaborative. Very few platforms are built with the understanding that both architectures need to be available depending on who's using the tool and what outcome they're after.
Consider the range of thought leadership work happening inside a typical mid-size law firm.
Marketing teams produce practice group newsletters, client alerts, award submissions, and case study write-ups. For most of that work, they're operating in delegation mode with raw material in hand, needing efficient execution.
Meanwhile, partners who want to establish themselves as thought leaders — writing articles for industry publications, building a LinkedIn presence, preparing keynote remarks — are doing something entirely different. Structured conversations that pull insights out of them produce their best content. Their workflow demands the collaborative mode.
Investing in a delegation-only tool means your marketing team will be happy while your partners get mediocre output because nobody asked them the right questions. Going all-in on collaboration means your partners might love the experience while your marketing team finds it painfully slow for routine content that should take 20 minutes.
At InnovAItion Partners, this is one of the hardest product questions we're wrestling with. We serve law firms, consulting firms, and marketing agencies — organizations where both modes of working exist under the same roof, sometimes within the same team.
Our approach, which is still evolving, is to build both capabilities into our AI assistants and let the user choose their mode based on the task at hand. Some days you want to paste your notes into a form and get a polished draft in three minutes. Other days you want to sit down with an AI that will interview you for 15 minutes and produce something you couldn't have written from a form.
Prompt chains handle the delegation side. Agentic frameworks like LangGraph handle collaboration. Both technologies exist today and are production-ready. Where it gets hard is in the design: making the experience intuitive enough that a marketing coordinator and a managing partner can both open the same tool and immediately find the mode that fits how they work.
I'll be honest — we haven't cracked it yet. And I'm skeptical of anyone who claims they have. Firms that recognize this as a design problem, rather than a "find the right AI tool" problem, are going to be significantly ahead of those still trying to force one mode on everyone.
So here's the question for your firm: do you know which mode each person on your team actually needs? Because chances are, the answer is different for almost every one of them.