Pharma manufacturing is generatingmore data than ever. The harder question is whether that data can actually helpsomeone make a better decision on the shop floor.

A batch record is digital.

The MES is digital.

The ERP is digital.

Laboratory systems are digital.

Equipment is producing sensor data.

Quality systems are producing deviation,CAPA and investigation data.

Supply chains are generating demand,inventory and logistics data.

Yet ask a manufacturing leader a simplequestion:

“Why did yesterday’s batch take 14%longer than expected?”

The answer often still requires people topull data from multiple systems, open spreadsheets, check equipment logs, speakto operators, review deviations and reconstruct what happened.

That is not a lack of data.

It is a lack of connected operationalintelligence.

And this distinction matters becausepharma manufacturing is now entering the next phase of digital transformation.

The question is no longer whetherpharmaceutical companies should use AI.

The question is whether theirmanufacturing data, processes and people are ready to make AI useful.

Theindustry is investing in AI. The factory still has a data problem.

The numbers show that the appetite for digital transformation isreal.

Deloitte’s 2025 Smart Manufacturing and Operations Survey of 600executives found that 92% of manufacturers believe smart manufacturing willbe a major driver of competitiveness over the next three years.

The same research found that manufacturers implementing smartmanufacturing initiatives reported average improvements of:

·        10–20% in production output

·        7–20% in employeeproductivity

·        10–15% in unlocked capacity

These are meaningful operational numbers.

But there is another number worth paying attention to.

Only 29% of surveyed manufacturers reported using AI or machinelearning at the facility or network level, while another 23% were stillpiloting AI/ML.

For generative AI, 24% reported deployment at facility or networklevel and 38% were still running pilots.

This tells us something important.

The manufacturing industry has moved beyond talking about digitaltransformation.

But AI adoption is still considerably less mature than the broaderdigital transformation conversation suggests.

And pharma has an additional constraint.

It cannot simply experiment with operational technology in the sameway an ordinary manufacturing business can.

Quality, validation, traceability, data integrity, patient safetyand regulatory expectations sit around almost every important manufacturingdecision.

That makes the underlying data problem more significant, not less.

The realproblem is not “too little data”

Walk into a modern pharmaceutical manufacturing organization and youare unlikely to find a shortage of data.

You may find:

ERP data

Materials, purchasing, inventory, production planning, costing andfinancial information.

MES data

Production execution, electronic batch records, work instructionsand process information.

LIMS data

Laboratory results, samples, tests and analytical information.

SCADA and historian data

Equipment parameters, temperatures, pressures, alarms and processconditions.

QMS data

Deviations, CAPAs, change controls, investigations and qualityevents.

Maintenance data

Equipment failures, work orders, preventive maintenance anddowntime.

Environmental monitoring data

Cleanroom conditions, microbial monitoring and environmentalobservations.

Supply-chain data

Demand, inventory, suppliers, transportation and lead times.

The problem is that these datasets were usually created fordifferent purposes.

They were designed to run a function.

Not necessarily to explain the entire manufacturing system.

That distinction becomes painful when management wants to answerquestions that cross organizational boundaries.

For example:

Why is batch cycle time increasing?

The answer may not live inside MES.

It could involve equipment downtime from the maintenance system.

A raw-material delay from ERP.

A repeated deviation from QMS.

An operator intervention recorded elsewhere.

A laboratory release delay from LIMS.

Or a process parameter drifting over several batches in thehistorian.

Each system may be working perfectly.

The organization can still struggle to understand what is happening.

This is the pharma data problem.

Digitalsystems do not automatically create digital intelligence

One of the most common mistakes in digital transformation isconfusing digitization with transformation.

Digitization means converting a manual activity into a digital one.

Digitalization means using digital systems to improve an existingprocess.

Digital transformation goes further.

It changes how the organization makes decisions and operates.

That difference is becoming increasingly relevant in pharmaceuticalmanufacturing.

A digital batch record is useful.

But a digital batch record that simply replaces paper is not thesame thing as an intelligent manufacturing system.

The first question is:

Can we capture the data?

The next question is:

Can we trust the data?

Then:

Can we connect it to other operational data?

Then:

Can we understand what it means?

And finally:

Can someone act on that insight fast enough to change the outcome?

AI only becomes useful near the end of this chain.

AI is nota shortcut around bad manufacturing data

This is where much of the current AI conversation becomesdisconnected from manufacturing reality.

A company can buy an advanced machine-learning platform.

It can deploy an LLM.

It can build a predictive-maintenance model.

It can create a digital twin.

It can connect a copilot to manufacturing documents.

None of those things automatically solve fragmented data.

Deloitte’s 2025 manufacturing research found that nearly 70% ofmanufacturers identified data-related issues—including data quality,contextualization and validation—as significant obstacles to AI implementation.The same research showed that 75% of respondents had increased investment indata lifecycle management to support GenAI strategies.

That is the less glamorous part of AI transformation.

Before asking:

“What AI model should we use?”

Pharma manufacturers should ask:

“Can we reliably explain what happened in our manufacturing processusing the data we already have?”

If the answer is no, AI will probably produce an impressivedemonstration rather than a dependable production capability.

Considera simple example: predicting batch failure

Imagine a manufacturer wants to predict the probability that a batchwill fail quality specifications.

The AI model could potentially use:

·        Process parameters

·        Raw-material characteristics

·        Equipment conditions

·        Historical batch outcomes

·        Environmental conditions

·        Operator interventions

·        Laboratory results

·        Deviations

·        Maintenance history

On paper, this sounds like an excellent AI use case.

But now consider the data.

The process data may be stored every few seconds.

The laboratory data may be recorded only at defined sample points.

The maintenance system may use equipment IDs that do not match theMES.

The deviation system may identify the batch differently.

Some operator interventions may be recorded manually.

Historical batches may have been manufactured using slightlydifferent procedures.

Some fields may have changed meaning over time.

Some records may be missing.

And some historical “failures” may have been corrected throughinvestigation and reprocessing.

The machine-learning problem is no longer simply:

“Can we build a good model?”

It becomes:

“Can we construct a trustworthy representation of the manufacturingprocess?”

That is a data-engineering problem before it becomes an AI problem.

The missing layer:context

Raw data rarely explains itself.

A temperature reading of 38°C means very little without context.

Which equipment?

Which batch?

Which stage?

Which product?

Which recipe?

What was the target range?

Was the equipment operating normally?

Was the measurement taken during cleaning, setup or production?

Was an operator intervention happening at the same time?

Was there a maintenance event?

Without context, an AI system can see numbers.

It cannot necessarily understand the manufacturing event.

This is why pharmaceutical manufacturers increasingly need a connectedoperational data layer rather than another isolated application.

The goal is not to create one enormous database for the sake ofhaving one.

The goal is to establish consistent relationships between:

Batch → Material → Equipment → Process Step → Operator Event →Quality Result → Deviation → Maintenance Event

Once those relationships exist, analytics becomes dramatically moreuseful.

And AI has something much better to work with.

Thesurprising gap: many plants already have the data they need

This is perhaps the most overlooked opportunity.

Pharma companies often assume that AI transformation requirescollecting huge amounts of new data.

Not always.

Sometimes the more valuable exercise is discovering what theorganization already has.

McKinsey’s 2025 analysis of biopharma production found that manyfacilities already collect high-quality process data and have data-collectioninfrastructure.

Yet among 20 biologics drug-substance manufacturing plants assessed,only 2 respondents—10%—reported using advanced analytics regularly forimprovement actions or real-time optimization.

That is a striking gap.

The factory is collecting data.

The organization is investing in infrastructure.

The information exists.

But the data is not consistently becoming operational decisions.

This is where digital transformation should become much morepragmatic.

The first objective should not be:

“Make the factory AI-powered.”

It should be:

“Make the factory explainable through data.”

AI can come after that.

Where AIcan actually create value in pharma manufacturing

Once the data foundation is strong enough, the opportunity becomesmuch more practical.

1. Predictive quality

Instead of waitinguntil the end of a process to discover that a batch has moved outside expectedconditions, machine-learning models can identify combinations of processconditions associated with quality outcomes.

The objective is notto replace quality decisions.

It is to givemanufacturing and quality teams earlier visibility.

That distinctionmatters in a regulated environment.

AI should help answer:

“Is this processbehaving differently from what we have historically observed?”

rather thanpretending:

“The model has decidedthe batch is good.”

2. Predictive maintenance

Equipment failurescan create much more than maintenance costs.

They can affect:

·        Production schedules

·        Batch cycle time

·        Capacity

·        Material utilization

·        Cleaning schedules

·        Quality investigations

·        Delivery commitments

Machine-learningmodels can combine equipment sensor data, alarms, maintenance history andoperating conditions to identify patterns associated with potential failures.

The value is notthe prediction itself.

The value ishaving enough lead time to act.

3. Process optimization

Manufacturing teamsoften operate within defined process windows.

AI and advancedanalytics can help identify which combinations of process parameters areassociated with stronger yield, shorter cycle times or more consistentoutcomes.

McKinsey estimatesthat GenAI could create $4–7 billion annually in value for biopharmaoperations through productivity improvements, workload reduction, equipmenteffectiveness and quality improvements.

The important wordis operations.

The opportunity isnot limited to chatbots.

4. Deviation and CAPAintelligence

Apharmaceutical company may accumulate years of:

·        Deviations

·        Investigations

·        CAPAs

·        Change controls

·        Audit observations

·        Equipment failures

·        Root-cause analyses

Much ofthis information is difficult to analyze collectively because it exists indocuments, structured records and different systems.

AI canhelp identify recurring themes.

Forexample:

Havesimilar deviations occurred across multiple plants?

Areapparently unrelated deviations connected to the same equipment family?

Arecertain raw materials associated with recurring investigations?

WhichCAPAs appear repeatedly without permanently eliminating the underlying problem?

This iswhere generative AI and traditional analytics can complement each other.

The LLMcan help retrieve and summarize evidence.

Analyticscan identify patterns.

Humanexperts remain responsible for the conclusion.

5. Manufacturingknowledge copilots

One of the more immediately practical GenAI opportunities is notautonomous manufacturing.

It is knowledge retrieval.

A manufacturing organization may have decades of knowledge spreadacross:

·        SOPs

·        Work instructions

·        Batch documentation

·        Investigation reports

·        Equipment manuals

·        Training documents

·        Quality procedures

·        Change controls

Finding the right information can consume significant time.

A properly governed AI assistant could help an authorized employeelocate relevant information faster.

But this is where pharma needs to resist the temptation to treat anLLM like an authority.

The system should show its sources.

It should respect access controls.

It should preserve auditability.

It should distinguish between approved procedures and generalinformation.

And it should not silently generate a new procedure and treat it asapproved.

The distinction between assistance and authorization isfundamental.

6. Supply-chainintelligence

Manufacturing performance does not start at the production line.

A delayed raw material can become a production problem.

A supplier-quality issue can become a deviation.

A forecast error can become excess inventory.

A capacity constraint can become a customer-service problem.

This makes supply-chain data another important layer ofpharmaceutical digital transformation.

AI can combine demand, inventory, supplier, manufacturing andlogistics signals to help identify emerging constraints.

The value comes from connecting information that was previouslyanalyzed separately.

Theregulatory question changes the architecture

There is another reason pharma AI transformation cannot simply copywhat consumer companies are doing.

In regulated environments, explainability, validation,traceability and human oversight matter.

The FDA’s January 2025 draft guidance on AI supporting regulatorydecision-making proposes a risk-based credibility assessment framework tiedto the model’s specific context of use.

That concept is useful beyond regulatory submissions.

A model should not simply be described as:

“95% accurate.”

The better questions are:

·        What is the intended use?

·        What data was used?

·        What population or process wasrepresented?

·        Where does the model performreliably?

·        Where does it fail?

·        What happens when confidence islow?

·        Who reviews the result?

·        What gets recorded?

·        When does a human override themodel?

This is particularly relevant because one of the sources you sharedhighlights a deeper issue in pharma AI adoption: workforce qualification.

The article points out that pharma has historically requireddocumented training and qualification for people operating equipment andvalidated processes, yet AI tools have often been introduced with much lighteronboarding. It also cites the industry’s growing concern about AI ROI nottranslating into shorter development timelines or better outcomes.

That observation should not be dismissed as an HR issue.

It is an operational-control issue.

The humanis still part of the AI system

This is one of the most important ideas for pharmaceutical leadersto consider.

An AI model does not operate in isolation.

Someone interprets the output.

Someone decides whether to act.

Someone may override it.

Someone records the decision.

Someone is accountable for the outcome.

That means the actual system is:

Data → Model → Human → Decision → Process → Outcome

Not:

Data → AI → Outcome

The difference is enormous.

The microbiome research article you shared makes a similar pointfrom a completely different scientific context. It argues that biologicalsystems are dynamic and complex, and that AI works best as a human-in-the-loopco-investigator, rather than as a replacement for scientific judgment. Italso highlights data standardization, population variability, bias,reproducibility and explainability as continuing barriers.

The same principle applies to manufacturing.

AI should strengthen the manufacturing system.

It should not become an unexplained decision layer sitting on top ofit.

Whatshould a pharma manufacturer do first?

Not every company needs a five-year AI transformation program.

A more practical approach is to start with the manufacturingquestions that already matter.

Step 1: Start with abusiness problem

Donot begin with:

“Wherecan we use GenAI?”

Beginwith:

“Whereare we losing time, capacity, quality or visibility?”

Examples:

·        Batch release delays

·        Unplanned downtime

·        Repeated deviations

·        Yield variability

·        Excessive investigation time

·        Production scheduling problems

·        Manual reporting

·        Raw-material variability

·        High operator dependency

Step 2: Map thedata behind the problem

Identifywhich systems contain the relevant information.

Forexample:

Batch cycle-time problem

MES+ ERP + maintenance + QMS + historian.

Nowthe problem becomes tangible.

Step 3: Measure datareadiness

Beforebuilding a model, test:

·        Completeness

·        Consistency

·        Accuracy

·        Timeliness

·        Lineage

·        Context

·        Historical availability

·        Access controls

This stepoften changes the AI roadmap.

Sometimesthe correct first investment is not an AI model.

It is dataintegration.

Step 4:Establish a common operational model

Theorganization needs consistent relationships between entities such as:

Plant

Line

Equipment

Material

Batch

Process

Parameter

Operator event

Quality event

Maintenance event

Thisis the foundation for analytics that can cross system boundaries.

Step 5: Choose onemeasurable use case

Agood first use case should have:

Ameasurable baseline

Availabledata

Aclear owner

Adefined decision

Ameasurable financial or operational outcome

Forexample:

Reduceunplanned equipment downtime by 15%.

isa better AI objective than:

Buildan AI-powered smart factory.

Step 6: Putgovernance around the model

Define:

·        Intended use

·        Model owner

·        Data owner

·        Validation approach

·        Human reviewer

·        Escalation rules

·        Audit trail

·        Monitoring

·        Retraining criteria

·        Change-control process

Themodel should become part of the operating process—not an experiment sittingbeside it.

Thefuture factory will not be the one with the most AI

This may be the most important conclusion.

The leading pharmaceutical manufacturer of the next decade may notbe the company with the most AI pilots.

It may be the company that can answer operational questions fasterand with greater confidence than its competitors.

Questions such as:

What changed?

Why did it change?

Which batches are at risk?

Which equipment is likely to fail?

Which deviations are repeating?

Where is capacity being lost?

Which process parameters matter most?

What should the operator investigate next?

What evidence supports that recommendation?

That is a very different definition of digital transformation.

It is not about replacing every existing system.

It is about connecting the systems that already exist so that theorganization can see the manufacturing process as one system.

Thepharma AI race may actually be a data race

The industry is likely to keep spending on AI.

It should.

There is substantial potential across discovery, clinicaldevelopment, manufacturing, quality and supply chain.

But the companies that create durable value will probably not be theones that simply deploy AI fastest.

They will be the ones that build the infrastructure around AI:

Trusted data.

Connected processes.

Clear context.

Defined decision rights.

Qualified users.

Governed models.

Measurable outcomes.

Deloitte’s research already shows that manufacturers are puttingsignificant attention into these foundations. Among surveyed manufacturers, 40%identified data analytics as a priority investment for the next 24 months,compared with 29% for AI and 27% for IIoT.

That ordering is revealing.

Data is not the work that happens before AI.

Data is part of the AI strategy.

A better question for pharma leadership

The conversation around AI inpharmaceutical manufacturing is often framed as:

“How can we use AI?”

A better question is:

“Which manufacturing decisions couldbecome materially better if our data were connected, trusted and available atthe right moment?”

That question changes theconversation.

It moves the discussion from technologyto operations.

From pilots to outcomes.

From models to decisions.

From dashboards to action.

And from “digital transformation” as anIT program to digital transformation as an operating capability.

The AI model may ultimately be the mostvisible part of the solution.

But it is rarely the part that determineswhether the solution works.

The real advantage starts muchearlier—when a pharmaceutical manufacturer turns fragmented operational datainto something its people can trust, understand and act on.

That is where AI starts becoming useful.

And that is where the next phase ofpharma manufacturing transformation will be won.

‍