The Screen/Room Distinction: The Only AI Framework You Need

There is one distinction that answers almost every practical question about what AI should handle in your company and what it should not.

The distinction is between screen work and room work.

Screen work is everything that happens on a screen before you enter the room: the report you compiled, the email you drafted, the briefing you assembled, the notification you formatted.

Room work is everything that happens in the room: the client relationship, the judgment call, the negotiation, the quality decision, the conversation that changes how someone thinks.

AI is very good at screen work. AI cannot do room work. The entire AI strategy for a $5M to $25M company flows from this distinction.

This article defines the screen/room distinction precisely, applies it to the primary work types at a non-tech mid-market company, and describes how the distinction resolves the practical decisions that more complicated frameworks require committees to make.

Defining screen work — with specifics by sector

Screen work defined

Screen work is the work that produces a structured output from available information, before a human uses that output to do room work. It has three characteristics:

• The inputs are available (the data, the information, the context)
• The output format is defined (the structure, the length, the vocabulary conventions)
• A human reviews the output before it reaches anyone who matters

The third characteristic is important: screen work always has a human review gate. The AI produces the draft. The human applies the room judgment before the output advances.

This is not a limitation: it is what makes AI-assisted screen work safe for regulated, client-facing, and quality-sensitive applications.

Screen work by sector

Manufacturing workflows:

• RFQ response drafting from specifications and capabilities matrix
• NCR documentation from defect data and inspection records
• Production schedule communications to customers from scheduling system data
• Management reporting from ERP data exports
• Supplier performance summaries from purchase order and delivery data

What manufacturing does not automate (room work): the quality manager’s determination that this batch’s NCR pattern suggests a systemic issue, the managing director’s negotiation about the delay timeline, the estimating lead’s judgment about whether this RFQ is worth pursuing given current capacity.

Distribution:

• Customer back-order notifications from exception data
• Account health summaries from CRM and order history
• RFQ responses from price lists and availability data
• Weekly operations briefing from system exports
• Supplier communications from purchase order status data

What distribution does not automate (room work): the account manager’s conversation with the at-risk commercial contractor who has been a customer for twelve years, the managing director’s read of the quarterly business review with the company’s largest account.

Healthcare:

• Payer appeal letters from denial reason codes and clinical documentation (with BAA governance)
• Care coordination communications from referral data
• Compliance report narratives from outcome data
• Staff communications from operations decisions
• Denial triage from payer data exports

What healthcare does not automate (room work): the physician’s clinical judgment about the appropriate treatment approach, the care coordinator’s conversation with the patient who is scared about the next step.

Professional services (engineering/legal/accounting):

• Work product first drafts from research notes and matter context
• Client status communications from matter tracking data
• Proposal sections from project portfolio and experience records
• Research synthesis from source materials

What professional services does not automate (room work): the attorney’s legal judgment about the case strategy, the engineer’s professional determination about the structural adequacy of the proposed design, the accountant’s judgment about the accounting treatment for a novel transaction.

Non-profit:

• Grant proposal narratives from programme logic and outcome data
• Funder reports from compliance tracking data
• Donor cultivation letters from giving history and programme highlights
• Board materials from operational summaries

What non-profit does not automate (room work): the executive director’s cultivation conversation with the major donor who is considering a transformational gift, the programme director’s judgment about whether the programme’s challenges are being described in a way the funder will accept.

Applying the distinction — the three practical decisions it resolves

Decision 1: Should this task be AI-assisted?

Apply the screen/room test: does this task have a screen component (structured output from available information with a defined format) that is distinct from a room component (human judgment, relationship presence, professional determination)?

Example applications:

“Should AI draft our customer delay notifications?” The notification has a screen component (structured communication from exception data in a defined format) and the account manager may add a room component (specific framing for a high-stakes relationship), but the notification itself is primarily screen. Yes: AI drafts, account manager reviews and personalises where the relationship requires it.

“Should AI conduct our client kickoff meetings?” The kickoff meeting is room work: establishing the relationship, reading the client’s unstated concerns, building the trust that the engagement depends on. No: AI is not appropriate for this task.

“Should AI produce our compliance reports?” The data compilation and narrative drafting are screen work. The characterisation of programme challenges in language the funder will accept is room work embedded in the screen document.

Partial: AI handles the structure and narrative drafting. The programme director handles the judgment passages.

Decision 2: Should AI produce a draft or a final output?

If the human adds room judgment to the output before it reaches the audience: AI produces the draft, human applies the judgment, human produces the final.

If the output requires no room judgment before it reaches the audience: AI produces the final, quality review gate confirms it meets standards.

The most common practical application:

Most routine customer communications (back-order notifications, status updates, appointment confirmations) require no room judgment.

The information is in the data, the format is defined, the tone standard is in the Foundation. AI produces the final. The team member reviews for accuracy before sending.

Some customer communications require room judgment: the communication with the client who is at the edge of a significant escalation, the communication that must address a relationship rupture.

AI produces the draft. The account manager applies the room judgment.

Decision 3: Is this task AI-appropriate at all?

If the task’s entire value resides in room presence and there is no screen component that produces useful information before the room interaction: AI is not appropriate.

If the task has a screen component (preparation, synthesis, documentation, follow-up) even if the primary value is room work: AI handles the screen component even if it cannot handle the room component.

Example: “Can AI help with our executive leadership coaching?”

The coaching session itself is room work. The coach’s presence, attunement, and judgment are the product. No AI component.

But: the preparation for the coaching session is screen work: reviewing the client’s prior session notes, synthesising the progress against development goals, identifying themes from the client’s recent written communications. AI can assist with the preparation. The coach does the session.

The screen/room framework applied to the AI system design

The Foundation documents are encoded room knowledge

Every element of the Foundation is a translation of room knowledge into screen specifications.

• The voice guide is the room knowledge of how this company’s people sound in conversation, translated into written form so AI can produce screen work that reflects that quality
• The communication standards are the room knowledge of how this company has learned to communicate with each customer tier, translated into specifications so AI can produce screen work at the right register for each audience
• The vocabulary guide is the room knowledge of what this industry’s work sounds like at quality, translated into vocabulary specifications so AI can produce screen work that sounds like a sector expert wrote it

The Foundation build is the process of encoding the company’s accumulated room expertise into a form that guides the AI’s screen production. This is why sector-specific Foundation building requires sector-specific practitioners: the room knowledge of a HVAC parts distributor is different from the room knowledge of a specialty CNC manufacturer.

The workflow selection is determined by where the screen/room boundary sits

The workflows that belong in the AI system are the ones where the screen component is the primary time cost and the room component is short or non-existent for most instances.

Proportion estimates:

The workflows where the screen component is 70% or more of the time cost belong in the AI system. The workflows where the room component is 70% or more of the value do not.

Common questions on the screen/room framework

”What about tasks that are fully screen work but high-stakes — like a major grant proposal section?”

High-stakes screen work belongs in the AI system with a correspondingly rigorous review gate. The major grant proposal section is screen work (structured narrative from programme logic and outcome data in a defined format) regardless of the stakes involved.

The stakes determine the depth of the review gate, not whether AI is involved.

The $2M grant proposal’s foundation narrative is reviewed more carefully than the $50,000 grant’s programme summary, but both are produced with AI assistance and reviewed by the programme director before submission.

High stakes is not a reason to avoid AI-assisted screen work. It is a reason to invest more in the Foundation quality and the review gate.

”What happens when AI gets good enough to do some room work — does the framework break?”

The framework shifts at the boundary rather than breaking.

If AI develops the capability to assess the emotional state of a client conversation in real time and produce communication guidance, that capability makes a previously room-only task partially screen-accessible. The framework still applies: the boundary just moves.

The current boundary (2026) is roughly: AI handles structured output production from available information.

AI cannot handle real-time relational attunement, safety-critical professional determination, or judgment that depends on uncodified contextual knowledge. These are the room work components that the framework identifies as outside the AI system today.

”Can the boundary between screen and room shift as the team develops AI fluency?”

Yes, in one specific direction. As the team develops AI fluency, they become better at providing the structured inputs that allow AI to handle a higher proportion of what was previously a mixed screen/room task.

The billing coordinator who has six months of AI fluency knows exactly how to specify the payer relationship context that allows AI to produce the appeal argument calibration as a first draft.

The room component does not disappear, but the coordinator’s fluency allows AI to handle a higher proportion of the room knowledge translation into screen specification.

The screen/room distinction also explains why what to automate first in your business always begins with the highest-frequency screen workflows — and why the restraint framework in restraint in AI strategy is essentially a set of rules for protecting room work from being automated prematurely.

Want to map your company’s primary workflows against the screen/room boundary?

The screen/room distinction is the only AI framework a $5M to $25M company actually needs because it answers every practical deployment question with a single test.

Every hour of screen work that AI reduces is an hour the human can spend in the room. That is the whole game.

Path one: map your top five workflows today. For each: identify the screen component (structured output from available information) and the room component (human judgment, relationship presence, professional determination). Estimate the time proportion of each. The workflows where the screen component is 70% or more of the time cost are the workflows that belong in the AI system.

Path two: bring in a partner. Phos AI Labs maps your company’s primary workflows against the screen/room boundary in the first Foundation build session, identifying which workflows AI handles, which humans own, and which are both. Thirty minutes, no deck. Start here.