Task by Task: The Workflows We're Handing to AI — One Decision at a Time | Lens Four by Sean Martin | Read by TAPE9

[♪ OUTRO MUSIC PLAYING [♪

Task by Task. The workflows were handing to AI one decision at a time by Sean

Martin, lens4 at SeanMartin.com. I look at the intersection of business,

technology, and messaging regularly through three lenses. How organizations operate

and run their programs. How innovation and market forces are reshaping what's

possible and how the language and narrative around technology shapes what gets

funded, prioritized, and trusted. This week all three lenses are pointing at

the same thing and the picture is clearer than most people are comfortable

admitting. Nobody decided to remove the human from the workflow. That's the

part worth sitting with. In boardrooms, in budget reviews, in vendor evaluations,

nobody stood up and said, let's build a business process with no humans in the

loop. What happened instead was a series of smaller decisions, each of them

reasonable, each of them local, each of them defensible. An HR team bought a

screening tool to handle application volume, a legal department licensed an AI

drafting platform to reduce outside counsel spend. A finance team deployed

automated invoice processing to close faster. None of those decisions on their

own looks like giving up control. But map them together, task by task across a

single workflow and something significant emerges. The human is already

optional across most of the process. That's what I want to examine this week. Not

whether AI should take on more of the work. That debate is largely settled in the

data. But whether organizations have consciously mapped what they've actually

handed over and what that means for how businesses operate, compete, and carry

accountability going forward. Let's start with the business operations lens. Are

we delegating efficiently or giving up control? The honest answer is both and most

organizations can't tell the difference yet. Let me trace two workflows that most

businesses run every week, not as edge cases, not as experiments, but as normal

operating processes with deployed tools and real outcomes. The first is the

hiring workflow. A job requisition opens. What happens next used to require a

recruiters judgment at every step? Here's what the same process looks like with

current tools. The first task is resume screening. Unilever reported saving over

1 million pounds annually after deploying AI screening tools across its

hiring pipeline. McDonald's rolled out paradoxes conversational AI called

Olivia across thousands of locations to handle applicant screening and

scheduling. Candidates move from application to interview without a human

recruiter touching the file. AI tools now rank and filter applicants in

seconds against criteria, a human set once, and that now runs autonomously at

scale. The second task is interview scheduling. A large US financial services firm

using good time reduced time to fill by weeks by automating calendar

coordination alone. The moment a candidate cleared screening and interview

invite went out within hours. No recruiter coordination required. The third

task is first round interviewing and assessment. Higher view used by JP Morgan,

Goldman Sachs, Amazon, Microsoft, Emirates Airlines, and hundreds of others

conducts asynchronous first round interviews with no human present. The candidate

records answers to preset questions on their own schedule. The AI analyzes speech,

language, and behavioral indicators and returns a ranked score. No recruiter

watches the recording until after the AI has already filtered and ranked the

pool. Emirates Airlines reduced its hiring cycle from 60 days to seven using

this approach. The human interviewer enters at round two, but by then the AI

has already determined who gets that meeting. The fourth task is offer

generation and outreach. Recruiting platforms including Lindy and recruiter

flows agent mode draft, personalize, and send offer communications and follow

up sequences autonomously. The offer letter is written before a recruiter opens

their inbox. The fifth task is onboarding initiation. End-to-end workflow

automation, deployable today through platforms like N8N, covers the full

pipeline from CV submission through assessment, scheduling, and status

tracking without human intervention at any step. Five tasks. Five separate

vendor decisions. Each one made independently, each one with its own ROI story.

And together, a process where a candidate can move from application to offer

without a single human making an active decision along the way. Now let's run

the same analysis across a legal department's standard contracting process.

The first task is matter intake and triage. Checkbox AI handles incoming legal

requests through intelligent chatbots that capture context, ask clarifying

questions, and route matters to the right team automatically. No paralegal

spending the morning clearing an email queue. The second task is legal

research. Harvey AI, now embedded in AM law, 100 firms, surfaces relevant

case law, statutes, and precedent across large document sets in minutes.

Lexus Plus AI provides contextual legal reasoning on demand. What used to be

a junior associate's full day is now a prompt. The third task is contract

drafting. Spellbook drafts contracts inside Microsoft Word. Legal on users

report NDA reviews dropping from two hours to 30 minutes. One managing

partner reported a 40% increase in billing capacity, not from doing

better work, but because AI wrote the first draft on every matter. The fourth

task is contract review and redlining. Luminance identifies anomalies and

flags deviations from playbooks across thousands of contracts simultaneously.

Kira extracts specific clauses at scale. Across the category, AI contract

review tools are reducing review time by 75 to 85%. A task that defined legal

practice for decades is compressing toward minutes. The fifth task is approval

routing and post execution management. Contract Poday handles routing,

obligation tracking, and compliance monitoring after signature. Ironclad

manages the full contract life cycle, renewals, explorations, obligation

triggers, without a paralegal maintaining a spreadsheet. Again, five tasks, five

products, five separate procurement decisions, and end-to-end, a contracting

workflow where a contract can move from request to executed agreement without a

lawyer authoring a single original clause, and this pattern runs across the

business, not just in these two functions. Finance has it. AI invoice processing

platforms, handle capture, validation, approval routing, and payment scheduling

end-to-end, with one hospital association reporting a reduction in batch

processing time from 10 hours to minutes. Customer service has it. Gartner

projects that agentic AI will resolve 80% of common customer service issues

without human intervention by 2029, up from effectively 0 in 2024. Security

operations has it, and I've had direct conversations with the vendors

building these tools. Edward Wu, founder of Dropzone AI, told me ahead of Black

Hat USA 2025. Nobody wants to be a Tier 1 analyst forever.

Subuguha of Stellar Cyber, described a digital army of AI agents that

filter 70 to 80% of alerts before a human analyst sees them. The pattern looks

the same whether the workflow is closing a contract or closing a security

incident. The business question this raises isn't whether the tools work, most of

them do. The question is whether organizations have a clear deliberate answer to

which tasks require a human decision, and why? Because right now many

organizations are answering that question by default, one purchase at a time

rather than by design. Gartner puts a number on the trajectory, at least 15% of

day-to-day work decisions will be made autonomously through agentic AI by 2028

up from essentially 0 in 2024. That is in a distant forecast. It's a projection

from a baseline that is already moving inside most mid to large enterprises

today. Now the innovation and market shifts lens. What is the market building

and how fast is it moving? The market knows exactly what it's building. It's not

naming it directly, but the architecture is unmistakable. Gartner predicts that

40% of enterprise applications will include integrated task-specific AI agents

by the end of 2026 up from less than 5% today. Not AI assistants that help people

do their jobs. Agents that do the job within defined parameters without waiting

for a human to initiate each step. By 2035, Gartner's best case scenario has

agentic AI driving approximately $450 billion in enterprise software revenue,

roughly 30% of the entire market. Notice what the market is selling, task

specific, not workflow replacing, not roll eliminating, task specific. One task at

the time, each one rationalized locally, each one with a discrete budget line and an

ROI model. The cumulative effect, a workflow that no longer requires human

participation, isn't what's being sold. It's what's being assembled. This is where

the business opportunity gets genuinely interesting and where the strategic gap

between leading and lagging organizations is opening up. The companies deploying

these tools aggressively are not doing so because they ran an experiment. They're

doing it because the economics are compelling in ways that compound over time.

Recruiter flow data shows recruiters saving six or more hours per week, a 33%

productivity increase per person. Legal on users report reducing outside

council dependency by thousands of dollars per contract cycle.

Individually, those numbers are meaningful. Across a workforce, across a fiscal

year, they represent a structural cost advantage that competitors without these

tools cannot match. But the more consequential shift isn't cost reduction. It's

speed and scale. A hiring process that moves at machine speed doesn't just save

money. It changes who gets the best candidates. A legal team that can review,

red line, and execute contracts in minutes rather than days doesn't just

reduce billable hours. It changes how fast the business can move on deals. The

organizations that figured this out early are already operating at a different

tempo than those still treating AI as a pilot program. Forester projects that

distributed AI workflows will capture 45% of enterprise workload capacity by

2026. That's not adoption at the margin. That's a structural shift in how

work gets done. The complication, and it's a real one, is that the vendor market is

significantly ahead of organizational readiness for what these tools actually

do. Gartner estimates that only around 130 of the thousands of companies now

claiming to offer agenteic AI are delivering genuine agenteic capability. The

rest are rebranding existing automation and RPA under a new label. A genuine

agenteic system reasons across tasks, adjusts based on outcomes, and handles

exceptions without a human rewriting the playbook. A rebranded chatbot executes

a fixed sequence and breaks at the edge case. Buying the ladder under the

belief it's the former is how organizations end up with expensive tools that

create new workflow fragility instead of removing old bottlenecks. The

cyber security sector is already several steps ahead of most enterprise

functions on this curve, and the outcomes are real, not experimental. As I

wrote in the first lens for article, the 72 minute gap, organizations deploying

agenteic SOC automation are realizing documented measurable budget savings. Dropzone

AI's Edward Wu described it plainly when we spoke at Black Hat USA 2025. At

roughly $36,000 per year, their platform ran 4,000 automated alert

investigations. A number that simply cannot be staffed at comparable cost.

Subo Guha of Stellar Cyber in two separate conversations with me at RSAC 2025

and Black Hat 2025 described their digital army of AI agents filtering 70 to

80% of incoming alerts, allowing analysts to focus on the fraction that

require human judgment. Both companies are emphatic that the value isn't

hypothetical. The savings are already in the operating budget. The market is

also generating the next layer of infrastructure, which is itself a leading

indicator of how far adoption has already gone. When AI agent identity

governance becomes a funded product category, and it has, it means organizations

have already deployed enough autonomous agents into production that they've

discovered they can't see what those agents are doing or control what systems

they can reach. Token security, named a finalist in the RSAC 2026 innovation

sandbox, was built entirely around this problem. Governing AI agent identities

with the same rigor applied to human users. Continuous discovery intent to

wear access controls, life cycle management from deployment through decommissioning.

Moderna has already scaled from 750 to more than 3,000 internal AI agents in a

single year. The governance market doesn't emerge until the adoption that

requires governance is already underway. That tells you where the actual

baseline is. Here is the structural shift worth watching closely. Right now,

organizations are assembling workflows task by task through separate vendor

decisions. The next phase of the market eliminates that friction entirely, and the

infrastructure for it is already being built. What's emerging is the

agentic orchestration platform. A single governed environment where workflows

can be defined in plain language, purpose-built agents can be selected,

configured, guard railed, and monitored, and the cumulative workflow is visible

as a designed hole rather than discovered after the fact as an accumulated

pile of vendor contracts. NINTEX, which serves more than 7,000 organizations

across 100 countries, announced its agentic business orchestration platform in

September 2025, explicitly positioning it as a single governed layer,

unifying legacy systems, manual processes, and AI agents. Their incoming agent

designer feature enables IT leaders and business technologists to build,

evaluate, and orchestrate specialized agents in a low-code environment

without writing code and without leaving the governance framework. IDCs

Maureen Fleming framed it directly. Agentic business orchestration represents a

shift toward coordinating people, systems, and AI agents in governed ways that

ensure automation and AI deliver measurable results at scale. Microsoft is

moving in the same direction at enterprise scale. Co-Pilot Studio, already

connected to more than 1,400 systems, allows agents to be built in natural

language, configured, monitored, and governed from a single interface. Every

agent now gets a Microsoft Entra Agent ID, an identity credential that enables

governance across the fleet. Microsoft's own framing for 2026 is pointed. The

transition is from AI that helps people do work faster to AI that handles work

on behalf of the organization, with humans escalating into exceptions rather

than executing by default. The pattern is the same whether you're watching

NINTEX, Microsoft, Salesforce, Agent Force, or ServiceNow's Agentic

capabilities. The market is converging on a platform model where the workflow

is defined up front in plain language. Agents are scoped to specific tasks with

explicit permissions. Guard rails are set before deployment rather than bolted

on after. Human oversight points are designed in rather than assumed, and the

full workflow is auditable and measurable as a system. The organizations that get

ahead of this transition will enter it with clear workflow maps and defined

accountability structures. The ones that don't will find themselves importing

their accumulated default choices into the new architecture and inheriting all

the governance gaps that came with them. Now the language and messaging lens.

Why does everyone say augment when the direction is replaced? Because augment

gets funded, replace gets scrutinized, and the actual outcome is somewhere

neither word honestly describes. There is a phrase that appears in virtually

every vendor pitch, analyst briefing, and enterprise communication about AI

and automation. We augment human capabilities. We don't replace them. It surfaces

in hiring tech. It surfaces in legal AI. It surfaces in financial

automation and customer service platforms. It is, at this point, essentially

obligatory in the category. The phrase is doing real work. It manages three

audiences simultaneously. Employees watching their job functions shift,

procurement committees answering to boards who want to see AI investment

justified, and regulators watching closely how AI is being deployed in

consequential decisions. Augment, not replace, threads all three needles

cleanly. It implies human oversight remains intact. Accountability structures

are unchanged, and the organization is being measured and responsible. But

walk the data back against that framing, and it doesn't hold up. Swim lane

projects AI will resolve or escalate over 90% of Tier 1 security alerts by

2026. Not assist with them. Resolve them. Gardner projects autonomous AI

handling 80% of customer service issues without human involvement by 2029.

The legal contract review tools marketing 75 to 85% time reduction aren't

augmenting lawyers. They're doing the task and asking the lawyer to review the

output. The hiring platforms aren't helping recruiters screen faster. They're

screening and asking the recruiter to validate the ranking. When the AI

handles 80% of the task and the human handles exceptions after the fact,

that's not augmentation in any meaningful operational sense. That's

oversight of an autonomous system. The distinction has direct implications

for where accountability lives, what skills the organization needs to

maintain, and what happens when the output is wrong. I explored a version of

this tension from an unexpected angle on the music evolves podcast in a

conversation with Chandler Lawn, AI innovation and law fellow at the

University of Texas School of Law, Drew Thurlow, adjunct professor at Berkeley

College of Music, and Puyia Parto-Navid, partner at Sayfarth Shaw. We were

talking about AI-generated music and who owns the output, but the

underlying question was the same one running through every enterprise workflow.

When the system produces the thing that used to require a human, what does

the human's role actually become? The music industry is a few years ahead of most

enterprise functions on this curve. Universal music group and Warner music

group both reached landmark settlements with AI music platforms in late

2025. The answers they're landing on involve drawing explicit lines around

what requires human creative judgment and what can be systematically produced.

Enterprise operations will need to draw the same kinds of lines, probably with

less drama, but with the same underlying logic. The language gap has a practical

consequence beyond messaging. When leadership describes every AI deployment

as augmentation, it becomes difficult to have honest internal conversations

about what the organization has actually delegated, where the accountability

gaps are, and what happens when a consequential decision turns out to be

wrong. That conversation is easier to have before the workflow is fully

assembled than after. Gartner's prediction that over 40% of

authentic AI projects will be canceled by end of 2027 is worth reading through

this lens. It's not because the technology fails. It's because organizations

bought capability without building the governance, accountability structures,

and organizational clarity to run it responsibly. The language that got the

tool funded made those harder conversations easier to avoid at purchase time.

They don't stay avoided. At Black Hat USA 2025, Marco Chapelli and I talked

after walking the floor about exactly this. When every vendor claims the same

positioning, the actual distinctions disappear from the buyer's view. In our post

show episode, we called it the marketing milkshake problem. Every vendor's

message going into the same promotional blender and coming out tasting the

same, regardless of what the underlying technology actually does. The agent

washing problem isn't just a market integrity issue. It's a decision quality

issue for every organization trying to figure out what they're actually

acquiring and what decision authority they're actually transferring. And now the

fourth lens. When did you decide to hand over control? And who was in the room

when you did? Here is what I keep coming back to when I look at all three lenses

together. And it's the thing I find myself saying in conversations that rarely

makes it into polished conference presentations. We are already past the point

of no return. The human optional workflow is not the exception being cautiously

piloted. It is the operational default for hiring, contracting, finance, customer

service, and security operations and organizations that made five individually

rational procurement decisions and never looked at what those decisions

assembled. That's not naivety. I want to be clear about that. The organizations

deploying these tools are not confused about what they're buying. What they

haven't done and what the vendors selling to them have never required them to do

is map the cumulative shape of those decisions before committing to them. And I

don't think that's an accident. When I look at the language the vendor market has

built around this transition, augment, not replace, human in the loop, AI

assisted, I don't read it as imprecise. I read it as precise in exactly the

right direction. These are companies staffed with product managers, lawyers and

communications teams who understand exactly what their tools do when deployed at

scale. Augment, not replace, threads every needle it needs to thread, employee

relations, procurement approval, regulatory scrutiny, board optics. It's not a

description. It's a strategy and it has worked because organizations bought the

framing along with the capability and now have workflows they couldn't honestly

describe as augmented with a straight face. So where does accountability land

when an AI assembled workflow produces a bad outcome? Right now, nowhere. That is

not hyperbole. The procurement signer approved a task specific tool with its own

contained ROI case. The vendor sold a product that performs as specified. The

workflow that those tools assembled collectively, that's in a gap between

contracts between org chart lines, between the legal definitions anyone drafted

when they wrote the terms of service. Nobody owns the workflow. Everybody owns a

task. That should be alarming. Not because bad outcomes are inevitable, but

because the accountability structure that would catch and correct a bad outcome

before it becomes a crisis. That structure doesn't exist yet in most

organizations. The efficiency gains are real and already in the budget. The

accountability architecture is still theoretical. What I'm watching closely is

whether the agentic orchestration platform changes this dynamic or accelerates

it. My honest read, both depending on the organization. A small group of mature

deliberate organizations, the ones who were already doing workflow mapping

before procurement, who already had security and legal at the design table, will

use NINTEX, co-pilot studio and platforms like them to do exactly what those

platforms were designed to enable. Define the workflow first, configure the

agents inside it, set the guard rails before deployment and maintain a complete

audit trail of what was delegated and why. For those organizations, the platform

genuinely forces the design conversation because you cannot configure guard

rails without deciding what you're guarding. For everyone else, the platform will

make accumulation faster and cheaper. The same five irrational decisions, each

locally rational, collectively unexamined, will just be easier to make in one

place. Here's the structural reality I think most organizations are not yet

reckoning with. The auditors haven't arrived yet. The regulatory frameworks that

will eventually require organizations to account for autonomous workflow

decisions, who authorize them under what criteria, with what human oversight and

how exceptions are handled, those frameworks are being drafted right now. GDPR

took years to land on AI. The EU AI Act is already in motion. The US regulatory

posture is slower but not absent. The window between we accumulated this

workflow through procurement, and we need to demonstrate we designed it with

intention is open, but it is not going to stay open. The organizations that use

that window to map what they've built, establish where accountability sits, and

make explicit decisions about what requires human judgment, not because the AI

can't do it, but because the organization has determined that accountability

requires a person will be positioned to operate without disruption when the

frameworks arrive. The ones that don't will discover that the workflow they

built by default is not the workflow they would have chosen under scrutiny. The

vendors knew what they were building. The buyers in most cases didn't ask the

right questions. The auditors haven't arrived yet. That window is closing. If this

analysis is useful to you, the full article with all references, data points, and

links to every podcast conversation mentioned is at seanmarton.com. Search for

lens4, or find me directly at seanmarton.com. And if these are the kinds of

conversations you want more of, the redefining cyber security podcast is where I

explore them in depth every week with the people building, buying, and breaking

these systems. You can find it wherever you listen to podcasts or at redefining

cyber security podcast dot com. Thanks for listening.