Two-Model Orchestration for Employee Advocacy Content

The problem with single-pass content generation

The standard approach to AI content generation is simple: give the model a prompt, get back a post.

For general-purpose writing, this works reasonably well. For employee advocacy — where the goal is to produce distinct, voice-matched posts for a CEO, a head of sales, an engineering manager, and three individual contributors from the same campaign brief — single-pass generation produces a specific failure mode.

Every post sounds like a variation of the same thing.

The model simultaneously decides what to say and how to say it. Without a structured planning step, it defaults to surface variation: change the opening line, adjust a few word choices, slightly shift the framing. The underlying argument stays the same because the differentiation logic never had its own dedicated step.

This is why Bloomberry uses a two-model orchestration architecture.

What two-model orchestration means

Two-model orchestration sequences two AI steps before a word of the final post is written:

Step 1: Planning model (fast, structured) A fast planning model reads the campaign brief — or signal, or ad-hoc prompt — and generates a structured content plan for each employee. The plan includes:

• The specific angle this employee should take
• The evidence or examples most relevant to their role and audience
• The format and length appropriate for their voice profile
• Any claim from the brief exclusion list that might surface for this employee

The plan is not the post. It is the spec for the post.

Step 2: Generation model (high quality, voice-aware) The generation model receives three inputs: the content plan, the employee's voice profile, and the platform target (LinkedIn or X). It writes the final post using the plan as its instruction set — not the raw campaign brief.

This separation is what produces genuine differentiation. The CEO plan says: open with a market-level observation, anchor in strategic implication, do not mention technical implementation. The Head of Sales plan says: open with a customer situation, anchor in how this changes the deal conversation, do not use product feature names without context. The generation model follows each plan separately.

The posts that come out are genuinely distinct — same campaign, different angles, different voices.

Why the planning step changes the output quality

The planning model's job is differentiation logic, not prose quality. It does not need to be the best writer in the room. It needs to be fast, structured, and accurate at mapping brief fields to employee context.

By giving differentiation its own model pass, the generation model can focus entirely on voice quality. It receives a specific, well-structured plan and a rich voice profile. Its only job is to write well.

Without the planning step, the generation model is trying to simultaneously:

• Read the full campaign brief
• Decide what angle each employee should take
• Apply each voice profile
• Produce well-written prose

Asking one model to do all of this in a single pass is asking it to compress work that benefits from being done in stages.

The claim exclusion layer

One of the more important things the planning model handles is claim routing. Campaign briefs in Bloomberry contain two claim lists: claims to include and claims to exclude.

The exclusion list is the governance layer — competitor disparagement, unverified statistics, off-brand language, pricing claims the company does not want attributed to employees on personal accounts.

The planning model evaluates each employee's natural angle against the exclusion list and flags any plan path that would likely require a generation model to produce an excluded claim. The generation model then receives a plan that already respects the exclusion constraints — rather than trying to filter the generation output after the fact.

Post-generation filtering catches surface-level matches. Pre-generation plan adjustment prevents the excluded content from appearing in the plan at all.

Voice profiles as the third input

Voice profiles are not prompts. They are persistent, structured representations of how each employee writes — built from their writing samples, their edits to generated posts, and their approved outputs over time.

When the generation model receives a plan, it does not receive instructions like "write in a conversational but professional tone." It receives a voice profile that has characterized this specific person's sentence length, hedging behavior, formatting preferences, phrasing patterns, and depth calibration.

The interaction between a specific content plan and a specific voice profile is what produces a post that reads like the employee wrote it — not like a content team wrote something in the direction of that employee.

What this means for campaign velocity

From a practical standpoint, two-model orchestration changes what a single campaign launch can produce.

Without it: marketing writes or commissions individual posts for each leader, which takes hours per person and does not scale past three or four people.

With it: marketing fills in a campaign brief, selects employees in scope, and generates LinkedIn and X drafts for every employee in a single run. The planning step runs fast. The generation step runs per employee. The entire run for a ten-person campaign typically completes in under two minutes.

The output still goes through an approval queue — marketing reviews, employees approve — but the content creation bottleneck is gone.

How this extends to signal-to-post

The same orchestration architecture applies to signal-to-post generation, not just campaign briefs.

When a signal surfaces in Bloomberry's signal intelligence feed — a competitor announcement, an industry development, a company milestone — the planning model maps the signal against the same employee roster. Each employee's plan reflects their role's natural angle on that signal. The generation model produces voice-matched posts from each plan.

The signal-to-post workflow is the reactive version of what campaigns handle proactively. Same architecture. Different starting point.

Why multi-model support matters

Bloomberry supports Claude Sonnet, GPT-4o, and Gemini Pro as generation models. The planning layer runs independently of the selected generation model.

This matters because no single model is best for every employee in every campaign. A generation model that excels at structured LinkedIn posts for executives may not produce the most natural-sounding content for a mid-level engineering manager. Multi-model support lets teams — and eventually, Bloomberry — select the generation model that best matches each employee's voice characteristics.

The planning layer is model-agnostic. It produces the same structured plan regardless of which generation model will consume it.

What the architecture does not change

Two-model orchestration does not change the human review requirement.

Every post generated through this pipeline — whether from a campaign brief or a signal — enters an approval queue before reaching the employee. Marketing reviews for brand alignment. Employees provide final approval before anything publishes.

The orchestration architecture improves draft quality. It does not remove the need for human governance before distribution.

That governance layer is not an afterthought in Bloomberry's design. It is the reason the platform can be used for employee content that appears on personal accounts. The approval workflow sits at the end of every generation path, regardless of how the content was created.

Further reading

• Employee advocacy platform architecture — the full module breakdown
• Employee advocacy software overview — category, buyer comparison, and how Bloomberry differs
• Advocacy campaign software — the campaign brief intake and generation flow
• Signal-to-post workflow — how signals become employee posts
• Campaign brief template — free 13-field template for structuring any advocacy campaign
• AI that learns your voice — how voice profiles are built and maintained